A method and system for learning multi-source attribution of interaction anomalies and intelligent regulation

Abstract

The application discloses a kind of learning interaction abnormality multi-source attribution and intelligent control method and system.This method obtains continuous learning interaction event stream, to form diagnostic observation unit with target knowledge dimension and task state vector;Respectively build content fixed and replace learner, learner fixed and replace equivalent content, and cross matching reference group of learner and content fixed and replace component, media or environment, calculate the direction residual signature of source direction;Separate the main source contribution of learner, teaching content, component, media, environment and the contribution between objects, and determine candidate source and confidence through cross-reference group consistency door;Generate single-source intervention instruction that keeps knowledge dimension, task state and non-target source unchanged and only changes candidate source, confirm, adjust or revoke attribution according to the decline of abnormal residual after intervention, and generate machine executable control instruction.The application can reduce the false attribution of learning abnormality and form a verifiable intelligent control closed loop.

Description

Technical Field

[0001] This invention relates to the fields of computer-aided learning, learning analytics, event stream processing, anomaly root cause analysis, and intelligent control. Specifically, it relates to a method and system for anchoring learning interaction events with knowledge dimensions and task state vectors, constructing cross-matching reference groups, calculating directional residual signatures, separating multi-source main effects from inter-object coupling effects, and generating machine-executable learning control instructions through single-source intervention verification. Background Technology

[0002] Digital learning systems can continuously record events such as answer results, response time, prompts, media playback, state transitions, page operations, and terminal environment. Existing learning state prediction and knowledge tracking methods typically estimate knowledge mastery based on learners' historical answer sequences; existing learning analytics systems can also collect learner behavior and evaluate online courses.

[0003] On the other hand, in educational data analysis, there are already latent factor models or mixed-effect models that estimate learner ability, question or content difficulty as different factors; general event flow anomaly detection and multivariate root cause analysis techniques can also calculate anomaly scores for continuous data and rank the importance of multiple input variables; adaptive learning systems can also recommend resources, learning paths or interactive controls based on learner data.

[0004] However, an anomaly in the learning process is usually influenced by the learner, teaching content, interactive components, media resources, operating environment, and combinations of these factors. Relying solely on response results or a single learner model makes it difficult to distinguish between learner deficiencies and deficiencies in content, media, or environment; using only general mixed-effects decomposition is easily affected by inconsistencies in knowledge points, task status, difficulty, and version; and simply outputting variable importance or anomaly scores cannot verify whether the identified sources truly caused the learning anomaly.

[0005] In particular, in controlled-state-driven learning experiments, learners may be in the same knowledge point but with different combinations of task states. If the reference data does not simultaneously constrain the target knowledge dimension and the task state vector, the resulting benchmark may mistake normal state differences for anomalies. Conversely, comparing multiple learners only within the same content fails to reveal coupling and adaptation issues between specific learners, specific content, and specific operating environments.

[0006] Therefore, a computer-implemented method is needed: to form comparable diagnostic observation units based on knowledge dimensions and task state vectors, to construct cross-matching reference groups along different directions such as learner-fixed, content-fixed, and learner-content co-fixed, to generate directional residual signatures pointing to the source, to distinguish between primary and coupled sources through cross-reference group consistency conditions, and to verify attribution through single-source intervention that only changes candidate sources while keeping other states unchanged, and finally to form control instructions that can be directly executed by the course operation module. Summary of the Invention

[0007] 1. In a continuous event stream involving multiple courses, versions, and terminals, establish comparable diagnostic observation units based on the target knowledge dimension and task state vector; 2. Avoid the common learner-content mixture effect model, which incorporates differences in task status, media, and operating environment into learner or content effects; 3. Differentiate the primary source contributions and inter-object coupling contributions of learners, teaching content, components, media, and environment through cross-reference directions; 4. Reduce erroneous attribution caused by single anomalies, sample bias, and baseline drift by ensuring cross-reference group directional consistency and evidence coverage conditions; 5. By maintaining the knowledge dimension and non-target source unchanged through single-source intervention, candidate sources are validated and the validation results are converted into machine-executable control instructions. Technical solution

[0008] To address the aforementioned technical problems, this invention provides a multi-source attribution and intelligent control method for learning interaction anomalies. Instead of directly attributing the original accuracy rate, this method anchors the target learning event to a combination of knowledge dimensions, task state vectors, and source objects, constructing cross-reference groups pointing to each source, and identifying the root cause based on whether the abnormal residual decreases after single-source intervention.

[0009] Receive continuous learning interaction events from multiple learning terminals. Perform event name unification, value encoding unification, timestamp alignment, duplicate event removal, and version verification. Write the target knowledge dimension, current task state vector, and learner, content node, component, media, environment, and version objects into the event index.

[0010] A diagnostic observation unit is formed using the target knowledge dimension, task state vector, source object combination, and time window. Events within the observation unit are sorted by event number and timestamp to form at least two diagnostic features from the following: correctness, response latency, prompt call, replay, invalid operation, state path deviation, task exit, and migration performance.

[0011] Construct at least three types of reference groups: learner reference groups with fixed content and state but replaced learners; content reference groups with fixed learners and states but replaced with equivalent content or versions; and reference groups with fixed learners, content, and state but replaced with components, media, or environments. Reference events also need to meet matching constraints regarding learning stage, prerequisite knowledge, task difficulty, time interval, and device capabilities.

[0012] Desired eigenvectors pointing to the source are generated for each of the three reference groups. The difference between the target eigenvector and each desired eigenvector constitutes the directional residual signature. Residuals in different directions have different attributional meanings, thus avoiding the use of a single overall benchmark to mask source differences.

[0013] Using constrained mixed-effects models, robust regression, probabilistic graphical models, matrix or tensor decomposition, or graph message passing, first estimate the main source contributions of learners, content, components, media, and environment; then remove the main source contributions from the residuals to calculate the coupling contributions of specific learner-content, learner-environment, or content-media combinations.

[0014] Candidate sources and attribution confidence are determined based on residual direction, cross-content or cross-learner reproducibility, evidence coverage, time window stability, and the separability of the highest and second-highest contributions across different reference groups. If a consensus threshold is not met, a definitive root cause is not output; instead, data collection or a supplementary control task is initiated.

[0015] For candidate sources that reach the intervention threshold, a single-source intervention instruction is generated. The instruction freezes the target knowledge dimension, task state vector, task difficulty range, and non-candidate source objects, and only replaces candidate source objects or adjusts their parameters. The system executes the intervention and collects subsequent events under the same state constraints.

[0016] Calculate the decrease in abnormal residuals before and after intervention. If the decrease reaches the confirmation threshold, a candidate source is confirmed, and a long-term control instruction is generated; if the confirmation threshold is not reached, the confidence level is reduced, a candidate source is switched, or data acquisition continues. The control instruction is directly parsed and executed by the learning experiment generator, task scheduler, component manager, media manager, or environment adapter.

[0017] Features and reference baselines are updated incrementally according to a sliding time window. If multiple source objects show overall changes in the same direction simultaneously and are unrelated to the intervention, it is identified as baseline drift and the reference group is updated; if the changes are concentrated only in candidate sources or specific combinations of objects, they are retained as anomalous contributions. Beneficial effects

[0018] 1. Anchor events using knowledge dimensions and task state vectors to reduce false alarms caused by incomparability between different task states; 2. By using three types of cross-matching reference groups, residual signatures pointing to learners, content, and presentation environment are generated respectively, thereby improving the source discrimination; 3. By first estimating the main source contribution and then calculating the interaction residuals, it is possible to identify anomalous coupling and adaptation between learners and content or environment; 4. By using cross-reference group consistency gates and confidence conditions, we avoid directly identifying single errors or differences in sample structure as root causes; 5. By freezing non-target sources and only changing candidate sources in single-source interventions, relevance attribution can be transformed into verifiable interventional attribution; 6. By confirming, adjusting, or withdrawing the attribution based on the decrease in residuals after intervention, a continuously updated diagnostic loop is formed; 7. Machine-executable control commands enable attribution results to directly drive learning experiments, content routing, component replacement, media switching, and environment adaptation; 8. By using event stream incremental updates and baseline drift identification, the interference of long-term data distribution changes on anomaly detection is reduced. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0020] Figure 2 This is a schematic diagram of the learning interaction event and diagnostic observation unit of the present invention.

[0021] Figure 3 This is a schematic diagram of the multi-source object association structure of the present invention.

[0022] Figure 4 This is a flowchart of the overall process of the method of the present invention.

[0023] Figure 5 This is a schematic diagram of the cross-matching reference group and directional residual signature of the present invention.

[0024] Figure 6 This is a schematic diagram illustrating the separation of the main source contribution and the coupling contribution between objects in this invention.

[0025] Figure 7 This is a schematic diagram of the cross-reference group consistency gate of the present invention.

[0026] Figure 8 This is a schematic diagram of the single-source intervention command structure of the present invention.

[0027] Figure 9 This is a schematic diagram of the closed loop of intervention verification and intelligent regulation in this invention.

[0028] Figure 10 This is a schematic diagram illustrating an embodiment of the present invention in collaboration with a learning experiment generation system and a teaching content internal defect localization system. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments are used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection. Those skilled in the art can use equivalent data structures, reference group generation algorithms, contribution estimation models, or control execution methods without departing from the essence of the present invention.

[0030] Task state vector: Structured data describing the controllable state dimensions and their values ​​in the current learning experiment, such as reference time, process perspective, completion association, role position or spatial relationship.

[0031] Source objects: learners, teaching content nodes, interactive components, media resources, operating environment and its version that may affect the learning interaction results.

[0032] Diagnostic observation unit: An event aggregation unit jointly defined by the target knowledge dimension, task state vector, source object combination, and time window.

[0033] Cross-matching reference group: A set of comparable events formed by replacing one source object with another, given a fixed portion of the source object and state conditions.

[0034] Directional residual signature: A set of residual vectors formed by the target diagnostic features relative to the expected features pointing to the reference group from different sources.

[0035] Main source contribution: An interpretable anomalous component that is stably generated by a single source object in multiple observation units.

[0036] Inter-object coupling contribution: After removing the single-source master contribution, the interaction anomalous component that persists only in a specific combination of source objects.

[0037] Consistency gate: a computational condition for determining whether a candidate root cause is verifiable based on the residual direction, recurrence range, and evidence coverage across reference directions.

[0038] Single-source intervention: Control operations that keep the target knowledge dimension, task state vector and non-candidate sources unchanged or within a preset tolerance, and only change the candidate source object or candidate source parameters.

[0039] See Figure 1The system includes an event access and state anchoring module, a cross-reference group construction module, a directional residual signature calculation module, a multi-source contribution and consistency determination module, a single-source intervention verification module, an intelligent control instruction generation module, and an execution module. The event database stores events, object versions, and intervention records. When the contribution of teaching content meets the conditions, the content node identifier and knowledge point identifier can be transmitted to an independent internal teaching content defect localization system. External intelligent models can be used for open answer feature extraction or strategy interpretation, but are not a necessary condition for implementing this invention.

[0040] See Figure 2 An event can be recorded using JSON, a protocol buffer, or a relational table. For example: { "learner_id":"L_8F3A", "kp_id":"aspect_completion", "content_node_id":"challenge_07", "component_id":"reverse_panel", "media_id":"voice_07", "runtime_env":"wechat_h5_ios", "version_id":"2.1.0", "task_state":{"time":"now","view":"process","link":1}, "event_type":"challenge_submit", "state_before":{"link":0}, "state_after":{"link":1}, "correct":false, "response_time_ms":8420, "hint_count":1, "audio_replay_count":2, "timestamp":1783824000123 } The system establishes diagnostic observation units based on knowledge dimensions, task state vectors, combinations of source objects, and time windows, avoiding the mixing of different state combinations under the same knowledge point. Diagnostic features may include error intensity E, standardized response latency T, prompt rate H, media replay rate R, invalid operation rate I, state path deviation P, migration loss G, and task exit rate Q.

[0041] See Figure 5 Let the target observation unit be O*=(L*,C*,U*,M*,E*,Z*,K*), where Z* is the task state vector and K* is the target knowledge dimension. The system constructs the following reference group: Learners are directed to a reference group R_L: C*, Z*, and K* are fixed, and components, media, and environment are kept within preset tolerances. Other learners whose learning stages and prior knowledge match are selected. The residuals of the target observations relative to R_L primarily characterize the deviation of the target learner from the same content group.

[0042] Content refers to the reference group R_C: With L*, Z*, and K* fixed, select the target learner's observations in equivalent content nodes or other content versions. The residuals of the target observations relative to R_C primarily characterize the deviation of the target content from its equivalent content for the learner.

[0043] The presentation environment points to the reference group R_E: L*, C*, Z*, and K* are fixed, while interactive components, media resources, or runtime environments are replaced. The residuals of the target observations relative to R_E primarily characterize the deviation from presentation or environmental factors.

[0044] Equivalent content nodes require the same target knowledge dimension, the same task state constraints, and task difficulty within a tolerance range, while avoiding direct reuse of surface markers that would reveal the answer. The matching distance can be calculated by weighting the differences in learning stages, prior knowledge, task difficulty, event time intervals, and device capabilities.

[0045] Let the target diagnostic feature be x*. For the reference event j in the source direction s, with feature x_sj and matching weight w_sj, the expected feature can be expressed as: x̂_s = Σ_j(w_sj·x_sj) / Σ_j w_sj.

[0046] The residual in the source direction s is r_s = x* - x̂_s. The system concatenates r_L, r_C, and r_E in a fixed-dimensional order to form the directional residual signature. For data with many outliers, the median, truncated mean, Huber estimation, or quantile regression can be used to generate the desired features.

[0047] See Figure 6 In one implementation, the residual of the nth diagnostic observation unit is modeled as follows: r_n = a_(L_n) + b_(C_n) + c_(U_n) + m_(M_n) + e_(E_n) + q_(L_n,C_n,E_n) + ε_n.

[0048] a, b, c, m, and e represent the main source contributions from learners, content, components, media, and environment, respectively; q represents the coupling contribution between objects; and ε represents the unexplained residuals. The solution can be obtained using a generalized linear mixed-effects model with regularization constraints, robust regression, probabilistic graphical models, or graph message passing.

[0049] Unlike the typical learner-question two-factor model, the input of this invention is not an unconstrained response matrix, but a multidimensional event residual anchored by the task state vector; and multiple residual signatures are generated using cross-reference groups with different fixed directions, and the coupling contribution between objects is calculated separately after the main effect is removed.

[0050] See Figure 7 Candidate root causes can only enter intervention verification if the contribution from at least one source exceeds the candidate threshold and the corresponding directional consistency condition is met. Learner sources need to be reproduced in multiple equivalent content sources, while the overall learners matching the target content are not abnormal; content sources need to be reproduced in the same direction in multiple matching learners, while the target learners are normal in equivalent content sources; environment or media sources need to show significant improvement after changing the environment or resources under the same learner, content, and state; coupling sources need to have insufficient individual main effects, while the interaction residuals of specific object combinations are stable and significant.

[0051] Attribution confidence can be calculated based on the number of valid events, reference group type coverage, cross-content or cross-learner reproducibility, sliding time window stability, and the separability between the highest and second-highest contributions. When evidence is insufficient, supplementary collection instructions are output instead of definitive labels.

[0052] See Figure 8 and Figure 9 A single-source intervention instruction should include at least the candidate source, the object to be changed, the frozen fields, the allowed tolerance, the intervention parameters, the execution conditions, and the termination conditions. For example, when verifying the content source, the learner, knowledge dimension, task state vector, difficulty range, components, media, and environment are frozen and replaced with equivalent content nodes; when verifying the environment source, the learner, content, and state are frozen, and only the runtime environment or resource loading method is switched.

[0053] Let the abnormal residual before intervention be r_pre, and the abnormal residual after intervention be r_post. The decrease in residual can be expressed as Δ = ||r_pre|| - ||r_post||, or as the relative decrease rate (||r_pre|| - ||r_post||) / (||r_pre|| + δ). When Δ reaches the confirmation threshold and the direction is consistent in the preset number of repeated interventions, the candidate source is confirmed; otherwise, the confidence level is reduced or other candidate sources are switched.

[0054] Once a candidate source is confirmed, the system generates machine-executable control commands. For example: { "diagnosis_id":"D_1024", "source_type":"learner_content_coupling", "target_id":["L_8F3A","challenge_07"], "knowledge_dimension":"aspect_completion", "action_type":"generate_single_variable_contrast", "frozen_fields":["knowledge_dimension","task_state.time","task_state.view","runtime_env"], "source_to_change":"task_state.link", "parameters":{"context_count":3,"hint_level":1,"modality":["visual","subtitle"]}, "guard_condition":{"confidence_min":0.68}, "termination":{"max_trials":3,"residual_drop_min":0.25} } When the learner source is confirmed, the system can generate supplementary learning or control experiments; when the content source is confirmed, it can route equivalent content, pause the target version, and call the content internal defect location interface; when the component, media, or environment source is confirmed, it can switch controls, resources, or runtime configurations; when the coupling source is confirmed, it can enable adaptation paths only for specific learner-content or learner-environment combinations.

[0055] The system can process event streams using sliding time windows. For new events, only the corresponding observation unit, reference group statistics, and contribution parameters are updated, avoiding the reprocessing of all historical data. If multiple contents, learners, and environments simultaneously exhibit changes in the same direction within a continuous window, and these changes are not related to any particular intervention, this is considered a baseline drift; the system updates the desired features after a stabilization window. If changes are concentrated in a single source or a specific combination of objects, they are retained as anomalies.

[0056] In the tense learning experiment, learners consistently misjudged "completion association" across different verbs, roles, and equivalent content nodes, but the matching learners generally performed correctly in the target content nodes. Learner-directed residuals and cross-content consistency both met the criteria. The system generated a univariate controlled experiment freezing the reference time and process perspective while only altering the completion association. After the intervention, the residuals for other knowledge dimensions remained unchanged, while the completion association residual decreased. The system confirmed the abnormality in this knowledge dimension on the learner's side and continued to generate targeted remedial learning experiments.

[0057] After the release of Challenge Content V3, multiple learners with varying skill levels exhibited similar anomalies in the same state dimension; these learners performed normally in equivalent Content V2 and other nodes. Content-directed residuals, group-wide consistency, and content contribution met the criteria. The system froze the knowledge dimension, task state, components, media, and environment, replacing V3 with the equivalent V2. After intervention, the anomalous residuals significantly decreased. The system identified the content source, paused V3, and invoked the internal defect localization interface of the teaching content to further pinpoint the question or interaction sequence.

[0058] The same learner and content functioned correctly in desktop browsers and Android WeChat environments, but in certain iOS WeChat environments, audio playback failed on the first attempt, and the replay rate and exit rate increased. The system switched to local audio and subtitle channels while maintaining the learner, content, knowledge dimension, and task status unchanged. After intervention, the residual decrease reached the confirmation threshold. The system then confirmed the environment and media source and continuously implemented a degradation strategy for the corresponding environment.

[0059] In the word order experiment, a certain group of learners performed normally in the presentation with a character screen and action arrows, but showed a significant increase in invalid operations and response latency in the presentation of purely abstract slots; other learners showed no significant difference between the two presentations. The main effects of learners and content alone were insufficient, and the learner-presentation coupling contribution reached a threshold. The system only switched the character screen and arrow presentation for this group of learners. After intervention, the residuals decreased, confirming the source of coupling, without labeling the learner or content as anomaly.

[0060] Event standardization, pseudonymization, and some feature calculations can be completed locally on the terminal. The server only receives the observation unit keys, features, and object version information required for attribution. The system does not use identity information such as name, face, or precise address as necessary attribution features, and does not form permanent learner labels based on a single event; open-ended answers or original recordings are processed separately after obtaining legal authorization.

[0061] This invention is not limited to English learning; it can be used for mathematical operation rules, scientific variable experiments, physical parameter experiments, reading evidence chains, and other learning activities that can be recorded by digital terminals. Cross-reference groups can be generated by rules, nearest neighbor retrieval, propensity scoring, representation learning, or other matching algorithms. The contribution model can run locally or have features or strategies provided by an external intelligent model, but the traceability relationship between the source object, residual signature, intervention object, and control output should be maintained.

Claims

1. A multi-source attribution and intelligent regulation method for learning interaction anomalies, executed by a computer, characterized in that, include: A continuous learning interaction event stream generated by multiple learning terminals is acquired. The events in the learning interaction event stream are associated with learner identifiers, knowledge point identifiers, teaching content node identifiers, interactive component identifiers, media resource identifiers, runtime environment identifiers, version identifiers, task status vectors, and timestamps. A target diagnostic observation unit is formed using the target knowledge dimension, the task status vector, the source object combination, and a time window. A diagnostic feature vector is generated from the target diagnostic observation unit. A cross-matching reference group is constructed for the target diagnostic observation unit. The cross-matching reference group includes at least: a learner-pointing reference group that replaces the learner while maintaining the same teaching content node, target knowledge dimension, and task status vector; and a learning reference group that maintains the same learning content node, target knowledge dimension, and task status vector. Under the condition that the learner, target knowledge dimension, and task state vector are the same, replace the content pointing reference group of the equivalent teaching content node or content version, and under the condition that the learner, teaching content node, target knowledge dimension, and task state vector are the same, replace the presentation environment pointing reference group of at least one of the interactive components, media resources, or runtime environment; generate corresponding expected feature vectors according to the learner pointing reference group, content pointing reference group, and presentation environment pointing reference group respectively, calculate the directional residual signature of the diagnostic feature vector relative to each expected feature vector, and separate the main source contribution of learner, teaching content, interactive components, media resources, and runtime environment, as well as at least one inter-object coupling contribution based on the directional residual signature; Based on the directional consistency condition, evidence coverage, and contribution separability across the cross-matching reference group, candidate anomaly sources, target knowledge dimensions, and attribution confidence are determined. When the attribution confidence reaches an intervention threshold, a single-source intervention instruction is generated. The single-source intervention instruction includes the frozen target knowledge dimension, task state vector, and non-candidate source objects, as well as candidate source objects or candidate source parameters to be changed. The single-source intervention instruction is then sent to at least one of the learning experiment generation module, task scheduling module, interactive component management module, media resource management module, or runtime environment adaptation module. Subsequent learning interaction events generated after executing the single-source intervention instruction are obtained, and the decrease in the anomaly residual after intervention relative to the anomaly residual before intervention is calculated. When the decrease reaches a confirmation threshold, the candidate anomaly source is confirmed, and a corresponding machine-executable control instruction is generated. When the decrease does not reach the confirmation threshold, the attribution confidence is reduced, the candidate anomaly source is switched, or a supplementary acquisition instruction is generated.

2. The method according to claim 1, characterized in that, Each learning interaction event also includes at least four of the following: session identifier, event number, pre-operation state, post-operation state, changed state dimension, learner prediction, actual system result, reverse state estimation, correctness, response latency, prompt call, media replay, invalid operation, task exit, and transfer result. The events are written into the target diagnostic observation unit after the field names are unified, the value encoding is unified, the timestamps are aligned, duplicate events are removed, and the version is verified.

3. The method according to claim 1, characterized in that, When constructing the cross-matching reference group, the matching distance is calculated based on at least three of the following: learning stage, prior knowledge status, task difficulty, event time interval, and device capability. Events with matching distances exceeding a threshold are eliminated, and the retained events are weighted according to their matching distances. The equivalent teaching content nodes have the same target knowledge dimension, the same task status constraints, and task difficulty within a preset tolerance range.

4. The method according to claim 1, characterized in that, The primary source contribution and inter-object coupling contribution are calculated using at least one of the following: constrained mixed-effects model, robust regression, probabilistic graphical model, matrix or tensor decomposition, or graph messaging. First, the primary source contribution of each source object is estimated based on the directional residual signature. Then, the primary source contribution is removed from the directional residual signature. Based on the remaining interaction residual, the inter-object coupling contribution between learners and instructional content, learners and operating environment, or instructional content and media resources is calculated.

5. The method according to claim 1, characterized in that, The directional consistency condition includes at least one of the following: the same learner exhibits anomalies in the same direction in multiple equivalent instructional content nodes and the matched learner does not exhibit group anomalies in the target instructional content node; Multiple matched learners exhibit anomalies in the same direction within the same instructional content node, while the target learner does not exhibit corresponding anomalies within the equivalent instructional content node; The corresponding abnormal decrease occurs when the same learner changes the interactive components, media resources, or operating environment in the same teaching content node; the contribution of each single main source does not reach the source threshold, while the interaction residual of a specific source object combination continues to reach the coupling threshold.

6. The method according to claim 1, characterized in that, The machine-executable control instructions include target object identifier, action type, control parameters, execution conditions, validity period, and termination conditions. When the learner source is confirmed, the action type includes generating an equivalent scenario experiment, a univariate control experiment, adjusting the open state dimension, task difficulty, or prompt level. When the source of the teaching content is confirmed, the action type includes routing to an equivalent content node, pausing the target version, or calling the internal defect localization interface of the teaching content. When the source of the interactive component, media resource, or runtime environment is confirmed, the action type includes replacing the interactive component, switching media resources, enabling a visual equivalence channel, reducing animation complexity, or changing the resource loading method. When the source of coupling between objects is identified, the action type includes generating an adaptation path for the corresponding learner and content or environment combination.

7. The method according to claim 1, characterized in that, The diagnostic feature vector and the expected feature vector are updated incrementally according to the sliding time window. When multiple source objects exhibit overall drift in the same direction in a continuous time window and the drift is unrelated to single-source intervention, the overall drift is used as the baseline drift and the cross-matching reference group is updated. When the drift is concentrated only on candidate source objects or a specific combination of objects, it is retained as an anomalous contribution.

8. A multi-source attribution and intelligent regulation system for abnormal learning interactions, characterized in that, include: The event access and state anchoring module is used to acquire continuous learning interaction event streams and form target diagnostic observation units and diagnostic feature vectors based on target knowledge dimensions, task state vectors, source object combinations and time windows; The cross-reference group building module is used to build learner-oriented reference groups, content-oriented reference groups, and presentation environment-oriented reference groups; The module includes a directional residual signature calculation module, used to generate corresponding expected feature vectors for each reference group and calculate directional residual signatures; a multi-source contribution and consistency determination module, used to separate the contributions of each main source and the coupling contributions between objects based on the directional residual signatures, and to determine candidate anomaly sources and attribution confidence based on cross-reference group directional consistency conditions, evidence coverage, and contribution separability; a single-source intervention verification module, used to generate single-source intervention instructions containing a frozen field and candidate sources to be changed, calculate the decrease in the abnormal residual after intervention relative to the abnormal residual before intervention, and confirm, adjust, or revoke the candidate anomaly source; and an intelligent control instruction generation module, used to generate and send machine-executable control instructions after the candidate anomaly source is confirmed.

9. An electronic device comprising a processor and a memory, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images