A cognitive trajectory learning method and system for discrete intervention

Abstract

The application relates to the technical field of cognitive tracking and discloses a cognitive tracking graph learning method and system for discrete intervention, which comprises the following steps: a learning terminal receives a discrete intervention operation through an input device, and a processor records a trigger time, a calculation completion time and a refreshing time of the same discrete intervention interaction record; when the refreshing time is later than the trigger time, a link closure value is generated according to the chronological relationship of the three time points and the relative position of the calculation completion time in a time window, and an effective gate value is generated; abnormal records are not sent to an interaction graph construction and do not participate in neighbor aggregation and supervision training. After the processor cleanses and encodes the discrete intervention interaction record, an interaction graph is constructed, a gated causal weight is obtained by adjusting a causal weight based on the effective gate value, then gated node representation values, gated prediction loss and gated training total loss are obtained, model parameters are updated and a prediction value is output.

Description

Technical Field

[0001] This invention relates to the field of cognitive tracking technology, and more specifically, to a cognitive tracking graph learning method and system for discrete interventions. Background Technology

[0002] In cognitive tracking applications, the learning terminal continuously receives discrete operations such as answering questions, requesting prompts, retrying, and selecting pages, and writes these operations into memory. The processor typically constructs an interaction graph based on questions, knowledge points, the order of occurrence, or a pre-defined sequence, and then uses the model to output cognitive status, mastery probability, or subsequent learning prompts. This process relies not only on the recorded content but also on the pairing stability between terminal input, processing, and display refresh.

[0003] In actual terminals, the entry of input events into the queue, the generation of screen data, and the display refresh callback are completed by different operational stages, which may be affected by task queuing, refresh cycle, and clock quantization. If the log system only matches based on business number, writing order, or event time, completed records and refresh records may be incorrectly classified into discrete intervention interaction records, and records near the boundary may also exhibit unstable pairings.

[0004] Once such records enter the interaction graph, they affect the connectivity between adjacent nodes and the weights of training samples. The processor may reuse untrusted data during neighbor aggregation and parameter updates, and the traceability of the interaction graph version, model parameters, and predictions stored in memory will also decrease. Existing cleaning processes mainly focus on missing fields, duplicate records, and response time normalization, but lack continuous discrimination of the execution order between input events, computation completion time, and refresh time, making it difficult to directly control the contribution of untrusted records to graph computation and supervised training on the terminal side. Summary of the Invention

[0005] This invention provides a cognitive tracking graph learning method and system for discrete interventions, which solves the technical problems mentioned in the background art.

[0006] To achieve the above-mentioned objectives, the present invention provides the following technical solution.

[0007] A cognitive tracking graph learning method for discrete interventions, applied to a learning terminal including an input device, processor, memory, and display device, comprising: The input device receives discrete intervention operations, and the processor records the trigger time, calculation completion time, and refresh time of the same discrete intervention interaction record; When the refresh time is later than the trigger time, the processor generates a link closure value based on the order of the three times and the relative position of the completion time in the time window formed between the trigger time and the refresh time. It also truncates the link closure values ​​that indicate physical execution order conflicts and retains the link closure values ​​that indicate normal physical execution order, thus generating a valid gate value. When the refresh time is no later than the trigger time or any time is missing, the processor will mark the corresponding discrete intervention interaction record as invalid and will not send it into the interaction graph construction or participate in neighbor aggregation and supervised training. The processor cleans and encodes discrete intervention interaction records to obtain node input values, constructs an interaction graph, and obtains causal weights. The processor adjusts the causal weights according to the valid thresholds corresponding to the connected nodes, so that the discrete intervention interaction records with unreliable physical execution order reduce the neighbor convergence contribution, thus obtaining the gated causal weights. The processor uses gated causal weights to perform neighbor aggregation to obtain gated node representations, and forms causal branches and association branches based on the gated node representations; The processor obtains the gated prediction loss by supervising training with effective threshold-weighted causal branches, merges the gated prediction loss and the uniform loss into the total gated training loss, updates the model parameters based on the total gated training loss, and outputs the predicted value.

[0008] Compared with the prior art, the present invention has the following substantial features and significant progress: This invention records the trigger time, completion time, and refresh time at the learning terminal, and integrates the link closure value and valid threshold value into the interaction graph processing, neighbor aggregation, and supervised training process. This reduces the impact of discrete intervention interaction records, such as callback pairing conflicts, near clock quantization boundaries, or missing time points, on causal weights, gated node representation values, and model parameter updates. The processor can reduce the propagation of duplicate messages and invalid training contributions caused by erroneous connections, and the memory synchronously retains the link closure value, valid threshold value, causal weight, gated causal weight, model parameters, and predicted values, facilitating the tracking of interaction graph versions and cognitive state outputs. Attached Figure Description

[0009] Figure 1 This is a system architecture and gating mechanism diagram of a cognitive tracking graph learning system for discrete intervention according to the present invention. Detailed Implementation

[0010] The following description is provided in conjunction with the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the invention; those skilled in the art can make equivalent substitutions or combinations for the specific implementations.

[0011] Example 1: The cognitive tracking graph learning method for discrete intervention described in this embodiment is applied to a learning terminal including an input device, a processor, a memory, and a display device. The learning terminal can be a smart learning terminal, a tablet computer, a personal computer, an online learning client, or an educational management terminal; as long as it can receive discrete intervention operations, record the time during the processing, save interactive data, and output predicted values, this embodiment can be implemented.

[0012] Specifically, the input, processing, and display stages in the learning terminal are typically completed asynchronously by event queues, processing tasks, and refresh callbacks. The actual physical execution order of the same discrete intervention interaction record may differ from the log writing order. If records are sent to the interaction graph solely based on business number or writing order, discrete intervention interaction records with unreliable physical execution order will form erroneous edges and be propagated as valid causal information during neighbor aggregation and supervised training. This implementation method forms link closure values ​​by using the trigger time, completion time, and refresh time, and then uses these link closure values ​​to form effective thresholds. These effective thresholds are then used for causal weights, gated causal weights, gated node representation values, gated prediction losses, and gated total training losses, thereby reducing the neighbor aggregation contribution of discrete intervention interaction records with unreliable physical execution order.

[0013] In one implementation, in the learning terminal 100, the input device 110 receives discrete intervention operations, which include answering questions, requesting prompts, retrying, and selecting pages.

[0014] Example 2: Before executing this implementation, processor 120 establishes storage space in memory 130 for storing raw logs, cleaned data, graph data, model parameters, and results. The raw logs store the time information generated by discrete intervention operations from input device 110 and existing callbacks from learning terminal 100. The cleaned data stores the discrete intervention interaction records after sorting, deduplication, missing data handling, and encoding. The graph data stores the interaction graph, nodes, edges, causal weights, gated causal weights, and gated node representations. The model parameters store the model parameters used in the gated graph learning module 160. The results store the predicted values, interaction graph version, and model parameter version.

[0015] Processor 120 assigns the same intervention record number to each discrete intervention interaction record. The same intervention record number is used to ensure that the trigger time, calculation completion time, and refresh time are written to the same discrete intervention interaction record. This same intervention record number can be an existing business number from learning terminal 100, or it can be an incremental number or a random unique number; this numbering method is only used to ensure that the same discrete intervention operation can be traced in the log.

[0016] After input device 110 receives discrete intervention operations, processor 120 converts the discrete intervention operations into discrete intervention interaction records. Each discrete intervention interaction record contains a content number, a response result, an intervention code, an event time, a sequential position, and the same intervention record number. The content number indicates a knowledge point, question, or interaction object. The response result indicates the learner's feedback to the corresponding content. The intervention code indicates a prompt, retry, page selection, or other discrete intervention operation. The event time is used to sort the discrete intervention interaction records. The sequential position maintains the relative order of the discrete intervention interaction records during the learning process.

[0017] When reading the status of input device 110, memory 130 and display device 140, processor 120 uses only the accessible interfaces already available in learning terminal 100.

[0018] Example 3: For the In the secondary discrete intervention, the processor 120 records the trigger time through the time acquisition module 150. After calculating the time and refresh time The trigger moment is the moment when the discrete intervention operation enters the event queue. The completion moment is the moment when processor 120 completes the corresponding screen data. The refresh moment is the moment when display device 140 refreshes the corresponding screen.

[0019] Triggering time Corresponding to the discrete intervention operation received by the input device 110. After the input device 110 receives the answer, prompt request, retry or page selection operation, the event queue of the learning terminal 100 generates a corresponding event, and the processor 120 reads the monotonic clock reading when the event enters the event queue and writes it into the same discrete intervention interaction record.

[0020] After calculating the time Corresponding to the screen data completed by the processor 120. After generating the corresponding screen data, updating the page status, or completing the pre-display processing according to the discrete intervention operation, the processor 120 reads the monotonic clock reading in the task completion callback and writes it into the same discrete intervention interaction record.

[0021] Refresh Time Corresponding to the refresh of the corresponding screen by the display device 140, the display device 140 generates a display refresh callback when actually refreshing the corresponding screen. The processor 120 reads the monotonic clock reading in the display refresh callback and writes it into the same discrete intervention interaction record.

[0022] When the refresh time is later than the trigger time, the processor 120 continues to calculate the link closure value and the valid threshold value. When the refresh time is not later than the trigger time, the processor 120 marks the corresponding discrete intervention interaction record as invalid. When any of the trigger time, calculation completion time, or refresh time is missing, the processor 120 marks the corresponding discrete intervention interaction record as invalid.

[0023] Example 4: The processor 120 observes the position of the completion time within a time window formed between the trigger time and the refresh time. This time window starts at the trigger time and ends at the refresh time. If the completion time is between the trigger time and the refresh time, it indicates that the physical execution order of input, processing, and display is consistent with the normal processing flow of discrete intervention operations; if the completion time is outside this time window, it indicates that the corresponding callback pairing conflicts with the physical execution order.

[0024] Link closure value Calculate using the following formula.

[0025] Triggering time Indicates the first The monotonic clock reading of the discrete intervention operation entering the event queue. This value is provided by the event callback corresponding to the discrete intervention operation received by input device 110.

[0026] After calculating the time This indicates that processor 120 has completed the first... The monotonic clock reading of the screen data corresponding to the discrete intervention is provided by the task completion callback of processor 120.

[0027] Refresh Time This indicates that the display device refreshes at 140. The monotonic clock reading of the screen corresponding to the discrete intervention. This value is provided by the display refresh callback of the display device 140.

[0028] When the physical execution order satisfies At that time, the link closure value It is positive, and the link closure value is... Not greater than When the calculated time is near the midpoint between the trigger time and the refresh time, the link closure value... near When the calculated time is close to the trigger time or refresh time, the link closure value... near When the calculated time is earlier than the trigger time or later than the refresh time, the link closure value... It is negative.

[0029] Link closure value The rationality of this can also be reflected by the following formula.

[0030] The above relationship illustrates that, under normal physical execution sequence, the link closure value... It also reflects whether the calculation completion time falls between the trigger and refresh times, and how close the calculation completion time is to the time window boundary. Processor 120 will then display the link closure value. Convert to valid threshold This causes the link closure values ​​that conflict with the physical execution order to be truncated, while the link closure values ​​with the normal physical execution order are retained.

[0031] Effective threshold By the Link closure value of discrete intervention interaction record The truncated value is used to control the extent to which the discrete intervention interaction record participates in subsequent graph learning processing. Effective threshold. for At that time, processor 120 does not allow corresponding discrete intervention interaction records to contribute to neighbor aggregation and supervised training. Effective threshold. Greater than At that time, processor 120 allows the corresponding discrete intervention interaction records to be recorded according to a valid threshold. The magnitude of the value is involved in the calculation of gating causality weights and gating prediction losses.

[0032] The link closure value and effective threshold, along with the corresponding discrete intervention interaction record, are stored together in memory 130. After the processor 120 saves the above data, it can trace the time basis of each discrete intervention interaction record during offline training, online cognitive tracking, or result verification.

[0033] Example 5: Processor 120 cleans and encodes discrete intervention interaction records to form fixed-length node input values. The cleaning process first sorts the discrete intervention interaction records according to event time. For discrete intervention interaction records with completely duplicate fields, processor 120 retains one record and deletes duplicate records. For discrete intervention interaction records that lack content numbers, response results, or intervention codes, processor 120 discards them or fills in the missing values ​​using preset methods.

[0034] For missing trigger time, calculation completion time, and refresh time, the processor 120 does not use preset missing values ​​to fill them, but instead marks the corresponding discrete intervention interaction record as invalid.

[0035] Processor 120 uses integer encoding or one-hot encoding for content numbers and intervention codes. Processor 120 uses binary encoding or multi-class encoding for response results. Processor 120 uses conventional normalization for response duration. Response duration can be obtained from the time difference between the input event and feedback event corresponding to the discrete intervention operation. Normalization methods can include max-min normalization, mean-variance standardization, or logarithmic transformation followed by normalization. The above encoding and normalization methods are conventional data preprocessing methods, used only to combine different fields into fixed-length node input values.

[0036] The node input value is denoted as Node input value Save the first Each node sends a numerical vector to processor 120. Node input values It includes at least a numerical field derived from the content number, response result, intervention code, response duration, and sequence position. The processor 120 writes the node input values ​​to the memory 130 for the gating graph learning module 160 to read from the memory 130.

[0037] Example 6: Processor 120 constructs an interaction graph based on the cleaned and encoded discrete intervention interaction records. Interactive diagram Stored in memory 130. Node set. It can be formed from knowledge points, questions, or discrete intervention interaction records. Edge set It can be formed by adjacent occurrences, common knowledge points, or pre-stored sequential relationships.

[0038] When knowledge points are used as nodes, the processor 120 aggregates discrete intervention interaction records under the same knowledge point into node input values ​​and forms edges based on the pre-stored sequential relationship or common question relationship between knowledge points.

[0039] When a question is used as a node, the processor 120 writes the answer, intervention code and response time corresponding to the question into the node input value, and forms edges based on adjacent occurrences or common knowledge points of the questions.

[0040] When discrete intervention interaction records are used as nodes, the processor 120 forms a node for each valid discrete intervention interaction record and forms edges based on adjacent occurrences in the event time.

[0041] Processor 120 writes the content number, response result, intervention code, response duration, and sequence position for each node. Processor 120 writes the connection type and original connection score for each edge. The connection type can be used to represent adjacent occurrences, common knowledge points, or pre-stored sequential relationships. The original connection score can be formed by the number of connections, adjacent relationship values, or pre-stored sequential relationship values, and a fixed source of original connection scores is maintained within the same interaction graph.

[0042] Processor 120 obtains causal weights Causal weight Represents a node For nodes The normalized contribution value. In a preferred implementation, processor 120 performs non-negative normalization within the neighbor set according to the original connection score to obtain the causal weight. Processor 120 stores the causal weight to verify the changes before and after gating processing.

[0043] When node Neighbor set When there is at least one neighbor node with an original connection score greater than zero, the processor 120 obtains the causal weight according to the following formula.

[0044] Original connection score Represents a node With nodes Unnormalized connection scores between nodes. Original connection scores. You can take the number of connections, the adjacency value, or the pre-stored sequential value, and use the same source in the same interaction graph.

[0045] when When the value is zero, processor 120 does not generate nodes. Causal weights pointing to neighboring nodes, preserving nodes Input values ​​to its own nodes and stop working on the nodes. Perform neighbor aggregation. This zero-denominator processing is consistent with the zero-normalization basis processing in the gating causal weight calculation.

[0046] When knowledge points are used as nodes, the original connection score The priority is to select the pre-stored sequential relationship values ​​between knowledge points or the number of connections between common questions.

[0047] When the question is used as a node, the original connection score The preferred selection is the number of adjacent occurrences or the number of connections between common knowledge points.

[0048] When discrete intervention interaction records are used as nodes, the original connection score The optimal value is the value of the adjacent time relationship of the events.

[0049] Example 7: Processor 120 adjusts the causal weights according to the valid thresholds corresponding to the connected nodes to obtain gated causal weights. For connected nodes... and nodes The edge, processor 120 will node effective threshold ,node effective threshold With causal weight They are used together for edge rights processing.

[0050] Gated Causal Weights This represents the node's contribution value after adding the valid thresholds at both ends of the node. Neighbor set. Represents nodes The set of directly connected nodes. Causal weight. Representing neighboring nodes For nodes The normalized contribution value.

[0051] When node When the corresponding normalization criterion is zero, the processor 120 retains the node's own node input value and stops performing neighbor aggregation on that node. Through this process, if all adjacent edges of a node cannot provide a valid contribution due to a valid gate value of zero, the processor 120 will not generate meaningless gated causal weights caused by zero denominators.

[0052] Processor 120 performs neighbor aggregation using gated causal weights to obtain the gated node representation value. This represents the node numerical vector obtained by using gating causal weights.

[0053] Node input value for 3D vector. Gated node representation value. for 1D vector. Self-transformation value for The training matrix. Neighbor transform values. for The training matrix.

[0054] When using multi-layer gated graph learning computation, the gated node representation value output by the previous layer is used as the node input value of the next layer. The self-transformation value and neighbor transformation value of the next layer are reset according to the dimension of the node input value of the next layer, so that the vector addition dimension of each layer is consistent.

[0055] Self-transformation value Used to process nodes node input values Neighbor Transformation Values Used to process neighbor nodes node input values .

[0056] function For common activation functions in graph neural networks, Sigmoid, ReLU, or similar functions can be used. The choice of these functions does not change the computational relationship between the effective gate value, the gated causal weights, and the representation values ​​of the gated nodes.

[0057] Example 8: Processor 120 forms causal and correlation branches based on the gated node representations. The causal branch is used to output the predicted value. The correlation branch is used to calculate the uniform loss. The preset retention ratio is denoted as... Preset retention ratio The range of values ​​is to Preset retention ratio The preferred value is .

[0058] Preset retention ratio The purpose is to use the information corresponding to higher gating causal weights for prediction, and to use the information that has not entered the causal branch to constrain the uniformity of the correlation output.

[0059] Causal branch output causal output value Causal output value Indicates causal branching on the first The sample belongs to the first The output probability is given by the cognitive state category.

[0060] Tag value Indicates the first The sample belongs to the first The class's label value. Number of samples. This indicates the number of records in a single training batch. (Number of categories) This indicates the number of cognitive state categories to be output. Processor 120 calculates the gating prediction loss based on the valid threshold.

[0061] For nodes Related candidate edge set The processor 120 sorts the data from high to low according to the gating causality weights and retains them according to a preset retention ratio. Calculate the number of retentions .

[0062] Candidate edge set Represents a node The set of edges surrounding the branch selection point. Number of edges to retain. This indicates the number of edges fed into the causal branch.

[0063] Processor 120 will sort before One edge and its corresponding node are sent to the causal branch, and the remaining edges and their corresponding gated node representations are sent to the association branch.

[0064] When the candidate edge set When empty, processor 120 reserves the node. The processor 120 uses its own gated node representation value and skips edge selection. When the gated causal weights of multiple edges are the same, the processor 120 sorts them according to the event time sequence of the corresponding discrete intervention interaction records; if the event times are still the same, they are sorted according to their sequential position.

[0065] lower limit of probability To avoid the input of the logarithmic term being zero, the lower probability limit is... The range of values ​​is to .

[0066] Boundary handling causal output values It is only used to calculate the logarithmic term in the gated prediction loss.

[0067] Gated Predictive Loss This represents the prediction loss weighted by the effective threshold. When the effective threshold of a discrete intervention interaction record is low, its contribution to the gated prediction loss decreases. When the effective threshold of a discrete intervention interaction record is zero, it is not included in the effective contribution of the gated prediction loss.

[0068] Related branch output related output value Associated output values Indicates the related branch to the first The sample belongs to the first The output probability is given by the cognitive state category.

[0069] uniform loss This is used to constrain the associated output values ​​to tend towards a uniform target, preventing the associated branches from carrying the main causal prediction information. In a typical implementation, uniform loss... The KL divergence between the correlated output values ​​and the uniform distribution can be used.

[0070] Boundary handling associated output values It is only used to calculate the logarithmic term in uniform loss.

[0071] Using the lower probability limit Even when the causal or correlated output values ​​are zero in floating-point calculations, the processor 120 can still stably calculate the gated prediction loss and the uniform loss.

[0072] Processor 120 combines the gated prediction loss and the uniform loss into a total gated training loss.

[0073] Total training loss for gating This represents the combined loss after incorporating the gated prediction loss.

[0074] Uniformity coefficient This represents the proportion of uniform loss in the total gated training loss. Uniformity coefficient. The range of values ​​is to The preferred approach is to determine whether the gated prediction loss no longer decreases or the prediction accuracy no longer improves based on the validation data.

[0075] If the sum of the valid gate values ​​of a training batch is zero, the processor 120 discards the batch and reads the next batch of data to avoid the denominator of the gated prediction loss being zero.

[0076] Processor 120 updates the model parameters based on the total gated training loss. Model parameters This represents all numerical parameters to be updated in the gated graph learning module 160. In one implementation using stochastic gradient descent, the model parameters are updated using the following formula.

[0077] Update step size This indicates the magnitude of the shift in a single parameter update. Update step size. The range of values ​​is to Processor 120 can also use Adam or similar optimizers to update model parameters.

[0078] After each training round, the processor 120 writes the model parameters, loss record, and round number into the memory 130.

[0079] When the gated prediction loss on the validation data no longer decreases for two consecutive rounds, the prediction accuracy no longer increases for two consecutive rounds, or the preset number of rounds is reached, the processor 120 stops updating and retains the model parameter file with better validation results.

[0080] Example 9: During the online cognitive tracking phase, after the learning terminal 100 acquires new discrete intervention interaction records, the processor 120 again records the trigger time, completion time, and refresh time, and calculates the link closure value and effective threshold value. The processor 120 generates node input values ​​using the same cleaning and encoding methods as in the offline training phase, and adds nodes and edges to the interaction graph according to the same connection rules.

[0081] Processor 120 calls the saved model parameters and performs gated graph learning computation on the new node and its neighbors. Processor 120 first obtains the valid gate values ​​corresponding to the new node and its neighbors, then obtains the gated causal weights based on the valid gate values ​​and causal weights, and then performs neighbor aggregation based on the gated causal weights to obtain the gated node representation values, and forms causal branches and association branches based on the gated node representation values.

[0082] The causal branch outputs the probability of each cognitive state category. Processor 120 selects the category with the highest probability as the predicted value. The predicted values, the same intervention record number, the model parameter version, and the interaction graph version are written to the memory 130. The display device 140 presents the predicted values ​​as a list, progress bar, or status indicator.

[0083] When content IDs or intervention codes not seen in the training phase appear during the online cognitive tracking phase, the processor 120 can fill them with preset unknown category codes or preset missing values. This processing is a routine data consistency process, the purpose of which is to ensure that the node input values ​​in the online cognitive tracking phase maintain the same field structure as the node input values ​​in the offline training phase.

[0084] As an optional implementation, after the input device 110 receives a question-answering operation, the trigger time is... First, the data is recorded; then, processor 120 completes the corresponding image data and records the completion time. Display device 140 refreshes the corresponding screen and records the refresh time. At this time, it is satisfied. Link closure value Positive, valid threshold Greater than The processor 120 feeds the discrete intervention interaction record into the interaction graph construction and participates in the calculation of the gated causal weights and gated prediction losses according to the effective threshold.

[0085] As an alternative implementation, the processor 120 completes the corresponding screen data very close to the trigger time or refresh time. In this case, although the physical execution order may satisfy... However, the link closure value near Valid threshold Also close The processor 120 does not treat the discrete intervention interaction record as a fully reliable record, but instead reduces its contribution to the gated causal weights and gated prediction loss. This mitigates the impact of unstable pairings near the clock quantization boundary on neighbor convergence and supervised training.

[0086] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A cognitive tracking graph learning method for discrete interventions, applied to a learning terminal including an input device, a processor, a memory, and a display device, characterized in that, include: The input device receives discrete intervention operations, and the processor records the trigger time, calculation completion time, and refresh time of the same discrete intervention interaction record; When the refresh time is later than the trigger time, the processor generates a link closure value based on the order of the three times and the relative position of the completion time in the time window formed between the trigger time and the refresh time. It also truncates the link closure values ​​that indicate physical execution order conflicts and retains the link closure values ​​that indicate normal physical execution order, thus generating a valid gate value. When the refresh time is no later than the trigger time or any time is missing, the processor will mark the corresponding discrete intervention interaction record as invalid and will not send it into the interaction graph construction or participate in neighbor aggregation and supervised training. The processor cleans and encodes discrete intervention interaction records to obtain node input values, constructs an interaction graph, and obtains causal weights. The processor adjusts the causal weights according to the valid thresholds corresponding to the connected nodes, so that the discrete intervention interaction records with unreliable physical execution order reduce the neighbor convergence contribution, thus obtaining the gated causal weights. The processor uses gated causal weights to perform neighbor aggregation to obtain gated node representations, and forms causal branches and association branches based on the gated node representations; The processor obtains the gated prediction loss by supervising training with effective threshold-weighted causal branches, merges the gated prediction loss and the uniform loss into the total gated training loss, updates the model parameters based on the total gated training loss, and outputs the predicted value.

2. The cognitive tracking graph learning method for discrete intervention according to claim 1, characterized in that, When the processor records the trigger time, calculation completion time, and refresh time, the trigger time is the time when the discrete intervention operation enters the event queue, the calculation completion time is the time when the processor completes the corresponding screen data, and the refresh time is the time when the display device refreshes the corresponding screen. Furthermore, the trigger time, calculation completion time, and refresh time are written into the same discrete intervention interaction record using the same intervention record number.

3. The cognitive tracking graph learning method for discrete intervention according to claim 1, characterized in that, When the processor generates the link closure value, it makes a judgment based on the position of the calculation completion time relative to the time window formed between the trigger time and the refresh time, so that the link closure value simultaneously reflects whether the calculation completion time is located between the trigger time and the refresh time and the degree to which the calculation completion time is close to the boundary of the time window.

4. The cognitive tracking graph learning method for discrete intervention according to claim 1, characterized in that, The processor saves the link closure value and the effective threshold value together with the corresponding discrete intervention interaction record, and saves the causal weight and the gating causal weight to verify the changes before and after the gating process.

5. The cognitive tracking graph learning method for discrete intervention according to claim 1, characterized in that, The processor cleans and encodes discrete intervention interaction records by: sorting the discrete intervention interaction records according to event time; discarding or filling in the missing values ​​for discrete intervention interaction records that lack content numbers, response results, or intervention codes; and encoding the content numbers, intervention codes, response results, and response durations to form fixed-length node input values.

6. The cognitive tracking graph learning method for discrete intervention according to claim 1, characterized in that, The processor constructs the interaction graph by: taking one of the following as nodes: knowledge points, questions, or discrete intervention interaction records; taking adjacent occurrences, common knowledge points, or pre-stored sequential relationships as edges; and writing content number, answer result, intervention code, response duration, and sequential position for each node.

7. The cognitive tracking graph learning method for discrete intervention according to claim 1, characterized in that, When the processor obtains the gated causal weights, it uses the valid gate values ​​at both ends of the node and the causal weights together for edge weight processing, and uses the weighted result of the valid gate values ​​and causal weights at the neighbor end as the normalization basis; when the normalization basis corresponding to the node is zero, the processor retains the node's own node input value and stops performing neighbor aggregation on the node.

8. The cognitive tracking graph learning method for discrete intervention according to claim 1, characterized in that, When the processor forms causal branches and correlation branches, it sends nodes and edges into the causal branches according to the order of gating causal weights and the preset retention ratio, and sends the remaining gating node representation values ​​and edges into the correlation branches. The processor uses the causal branch to output predicted values ​​and the correlation branch to calculate uniform loss.

9. A cognitive tracking graph learning system for discrete interventions, characterized in that, The system includes a learning terminal, which comprises an input device, a processor, a memory, and a display device; the input device is used to receive discrete intervention operations; the memory is used to store discrete intervention interaction records, interaction diagrams, model parameters, and predicted values. Display devices are used to provide refresh times and output screen data; The processor is used to execute a cognitive tracking graph learning method for discrete interventions as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed by a processor, implement the cognitive tracking graph learning method for discrete intervention as described in any one of claims 1 to 8.

Images