Learning strategy evaluation and guidance method based on experiment operation sequence pattern mining

Abstract

The application discloses a learning strategy evaluation and guidance method based on experimental operation sequence pattern mining. The method collects original operation logs and reconstructs them into normalized operation event sequences; extracts the silent interval features between adjacent operations, analyzes the cognitive state labels corresponding to the silent interval in combination with the operation context, embeds them into the sequence to generate semantic enhanced operation sequences; then identifies the repeated operation fragments in the sequence, calculates the exploration intention features to distinguish the scientific exploration intention from the random trial intention, extracts the strategy feature vector of the student in combination with the typical strategy mode library; further locates the key decision branch point, extracts the decision micro-behavior features to calculate the decision hesitation index, and evaluates the strategy efficiency; and plans the gradual transition path from the current strategy to the target strategy in combination with the real-time cognitive load level, and generates personalized guidance suggestions. The application can realize deep quantitative analysis and accurate guidance on the cognitive state, operation intention and decision psychology of students.

Description

Technical Field

[0001] This invention relates to the field of educational data mining and intelligent tutoring technology, and in particular to a method for evaluating and guiding learning strategies based on experimental operation sequence pattern mining. Background Technology

[0002] In virtual simulation experimental teaching, in-depth analysis of students' inquiry processes helps to achieve intelligent teaching. By deeply mining the massive amounts of operation log data generated by students on the experimental platform, we can reconstruct students' cognitive paths at a micro level and quantitatively analyze their thinking strategies in hypothesis testing, variable control, and result reflection. This technology can break through the limitations of traditional teaching that only focuses on experimental results, and build an adaptive intelligent learning system (ITS) to achieve real-time diagnosis and intervention of students' learning process.

[0003] Currently, experimental teaching assessment technologies focus on rule-based result determination and superficial behavioral statistical analysis. Existing solutions retrieve students' operation logs from the system backend and statistically analyze the completion rate of their operation steps, the accuracy rate of experimental conclusions, and the total time taken to complete the task. In terms of process analysis, existing technologies mostly use traditional sequence pattern mining algorithms (such as the PrefixSpan algorithm or the Apriori association rule mining algorithm) to extract high-frequency operation subsequences, or simply divide the time interval between operations into long pauses and short pauses by setting a globally fixed time threshold, using this as a basis for segmenting operation tasks or judging whether students are active.

[0004] However, existing technologies primarily treat operation logs as discrete sequences of symbols, neglecting the deep cognitive semantics implicit in operational gaps and micro-behaviors. This results in coarse-grained policy evaluation and inaccurate intent recognition. Therefore, further research and innovation are needed to address these issues in existing technologies. Summary of the Invention

[0005] Purpose of the invention: In view of the above-mentioned problems in the prior art, this application provides a learning strategy evaluation and guidance method based on experimental operation sequence pattern mining.

[0006] Technical solution: According to one aspect of this application, a method for evaluating and guiding learning strategies based on experimental operation sequence pattern mining includes:

[0007] Collect raw operation log data generated by students on the experimental platform and reconstruct it into a standardized sequence of operation events based on the operation timestamps;

[0008] Extract the silence interval features between adjacent operations in the normalized operation event sequence, combine them with operation context features to parse the cognitive state labels corresponding to the silence intervals, and embed them into the normalized operation event sequence to generate semantically enhanced operation sequences.

[0009] Identify repetitive operation segments in semantically enhanced operation sequences, calculate exploration intention features to distinguish operation intention types, and extract students' strategy feature vectors by combining typical strategy pattern libraries;

[0010] Locate key decision branch points in semantically enhanced operation sequences, extract decision micro-behavioral features during the decision-making process, calculate decision hesitation index, and combine with policy feature vectors to generate policy effectiveness evaluation results;

[0011] Real-time cognitive load levels are calculated based on semantically enhanced operation sequences. Combined with policy effectiveness evaluation results, a gradual transition path from the current policy to the target policy is planned, and personalized policy guidance suggestions are output.

[0012] Beneficial effects: This invention can achieve in-depth quantitative analysis and accurate guidance of students' cognitive state, operational intentions, and decision-making psychology. The related technical effects will be described in detail below with reference to specific embodiments. Attached Figure Description

[0013] Figure 1 A flowchart illustrating a learning strategy evaluation and guidance method based on experimental operation sequence pattern mining, provided for embodiments of this application.

[0014] Figure 2 This is a flowchart illustrating the identification of repetitive operation segments in a semantically enhanced operation sequence, as provided in an embodiment of this application.

[0015] Figure 3 A flowchart illustrating how computational exploration intent features distinguish operational intent types in embodiments of this application.

[0016] Figure 4 This is a flowchart illustrating the extraction of student strategy feature vectors using a library of typical strategy patterns, provided as an embodiment of this application.

[0017] Figure 5 A flowchart illustrating the micro-behavioral characteristics of the decision-making process provided in this application embodiment. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] To address the aforementioned issues, the applicant conducted in-depth searches and analyses, and discovered:

[0021] Existing time analysis methods based on fixed thresholds cannot distinguish between purposeful thinking and confusion / stagnation in context, leading to misjudgments of students' cognitive states. Sequence mining based solely on operational symbols is insufficient to quantify the regularity in parameter adjustment processes and cannot differentiate between scientific exploration and random trial, two types of operational intentions that are similar in external form but differ in cognitive depth.

[0022] Furthermore, the lack of quantitative means for measuring micro-behaviors at decision points (such as repeated cancellation and preview) makes it difficult to pinpoint students' cognitive hesitation, causing subsequent guidance strategies to fail due to a lack of relevance or excessive cognitive load.

[0023] To solve these problems, combined with Figures 1 to 5 The present invention will be specifically described through the following embodiments.

[0024] On the one hand, this paper provides an exemplary scheme for evaluating and guiding learning strategies based on experimental operation sequence pattern mining. This solves the technical problems of low utilization rate of operation log data and difficulty in reconstructing students' cognitive processes from raw data in existing experimental teaching. Specifically, it includes:

[0025] Step 101: Collect the raw operation log data generated by students on the experimental platform and reconstruct it into a standardized operation event sequence based on the operation timestamp.

[0026] In this embodiment, the raw operation log data is the complete behavioral trace automatically recorded by the background database when students interact with the virtual simulation experiment system or online experiment platform. Specifically, the system exports raw records from the database containing fields such as user identifier, operation timestamp, operation type code, operation object identifier, operation parameter set, and system response status. To address potential differences in field naming across different experimental modules, a unified field mapping rule table is established to convert the logs from each module into a standardized recording format.

[0027] Furthermore, in the process of reconstructing into a standardized sequence of operation events, the original data needs to be cleaned and its temporal sequence reconstructed. Specifically, for continuous repetitive operation records generated by the same user within a very short period of time, the system determines that they are accidental touches or system jitter and performs debouncing processing.

[0028] As an optional implementation, the system sets a time threshold of 200 milliseconds, meaning it removes duplicate operation records with a time difference of less than 200 milliseconds from the previous record, retaining only the first record. For records with missing timestamps, linear interpolation is used to complete the timestamps based on the time difference between adjacent records, ensuring the continuity of the time sequence. After cleaning, the system abstracts each valid operation record into an operation event tuple E, whose physical structure can be represented as E = <Type,Obj,Params,t _abs , t _rel > where Type represents the operation type, Obj represents the operation object, Params represents the operation parameter vector, and t _abs Represents an absolute timestamp, t _rel This represents the relative time offset from the start point of the sequence. By arranging the above tuples in ascending order of timestamps, an independent chain of operation events, i.e., a normalized sequence of operation events, is constructed, with each student as a unit.

[0029] Step 102: Extract the silence interval features between adjacent operations in the normalized operation event sequence, combine the operation context features to parse the cognitive state labels corresponding to the silence intervals, and embed them (cognitive state labels) into the normalized operation event sequence to generate a semantically enhanced operation sequence.

[0030] Correspondingly, the silence interval is not an ineffective blank period, but rather a carrier for students' cognitive activities. The system traverses the normalized sequence of operational events, calculating the time difference between adjacent operational events as the fundamental physical quantity for the silence interval feature. Furthermore, the system extracts local contextual information to analyze the cognitive state the student may be in during this interval, such as purposeful thinking, confusion / stagnation, distraction, or system waiting. Implicit temporal information is made explicit as semantically meaningful labels, inserted as new nodes into the original sequence, thus achieving semantic enhancement of the operational sequence.

[0031] Step 103: Identify repetitive operation segments in semantically enhanced operation sequences, calculate exploration intent features to distinguish operation intent types, and extract students' strategy feature vectors by combining typical strategy pattern libraries.

[0032] In this embodiment, the system uses a sliding window algorithm to identify consecutively occurring subsequences of the same type of operation, i.e., repetitive operation segments. To distinguish whether students are engaging in purposeful scientific exploration (such as the controlled variable method) or aimless random trial and error, the system calculates exploration intent features, including the regularity of parameter changes, observation behavior, and coverage, and assigns intent labels accordingly. Simultaneously, the system introduces a pre-built library of typical strategy patterns as a reference, matching students' operation sequences with high-performance patterns in the library to generate a strategy feature vector describing the student's strategy type.

[0033] Step 104: Locate the key decision branch points in the semantically enhanced operation sequence, extract the micro-behavioral features of the decision process to calculate the decision hesitation index, and combine the strategy feature vector to generate the strategy effectiveness evaluation result.

[0034] Accordingly, the system focuses on key nodes in the experimental task, namely critical decision-making branch points. At these nodes, students face multiple path choices, and their behavioral characteristics can reflect their cognitive state. The system extracts micro-behaviors within the decision window, such as repeatedly viewing options (preview), selecting and then canceling (cancel and rollback), and switching back and forth between options (switching frequency), and combines this with the analyzed cognitive state to quantify the decision hesitation index. Based on this, the system integrates the matching degree of the strategy, the smoothness of execution (reflected by hesitation), and the effectiveness of the output to generate a multi-dimensional strategy effectiveness evaluation result.

[0035] Step 105: Calculate the real-time cognitive load level based on the semantically enhanced operation sequence, combine the policy effectiveness evaluation results to plan a gradual transition path from the current policy to the target policy, and output personalized policy guidance suggestions.

[0036] In this embodiment, the system monitors students' cognitive load in real time, avoiding forced intervention when students experience cognitive overload (such as frequent errors or prolonged confusion). Based on the evaluation results, the system plans a progressive path for students from their current inefficient strategies to a target efficient strategy. This path does not directly provide a standard answer but is broken down into several moderately challenging improvement stages, allowing students to gradually optimize their strategies within a manageable cognitive load. Based on this, the system outputs personalized guidance information containing specific operational suggestions or thinking prompts.

[0037] On the other hand, this paper describes an alternative implementation of temporally enhanced cognitive state parsing, particularly utilizing operational context features and a personalized temporal baseline model to parse silent periods into cognitive state labels with explicit semantics, thereby constructing semantically enhanced operational sequences. Accordingly, this embodiment can be implemented through the following steps:

[0038] Step 201: Calculate the timestamp difference between adjacent operation events in the normalized operation event sequence as the silence interval duration; extract the operation context features corresponding to each silence interval, which include: the complexity level of the preceding operation, the type conversion relationship between the subsequent operation and the preceding operation, and the relative position ratio of the current silence interval in the sequence; wherein, the type conversion relationship includes continuation of the same type, cross-type switching and reverse rollback. The silence interval feature is a composite feature that integrates the silence interval duration and the operation context features.

[0039] Furthermore, the system calculates the basic time interval. To accurately understand the meaning of this interval, the system extracts its context. The complexity level of preceding operations is determined based on a predefined knowledge base; for example, viewing data has low complexity, while configuring multi-level parameters has high complexity. Higher complexity corresponds to a longer expected reasonable thinking time. Type conversion relationships reflect the logical flow of operations: if subsequent operations belong to the same functional module as preceding operations and parameters are fine-tuned, it is considered a continuation of the same type; if it jumps to a different functional module, it is considered a cross-class switch; if it executes undo or return, it is considered a reverse rollback. Contextual features provide a logical basis for subsequent judgments.

[0040] Step 202: Perform distribution fitting on the silence interval duration in the normalized operation event sequence to construct an individualized time distribution baseline model based on a Gaussian mixture model. The individualized time distribution baseline model contains multiple mixture components corresponding to different operation rhythms. Calculate the standardized deviation of the current silence interval duration relative to each mixture component in the individualized time distribution baseline model. When the standardized deviation is less than a preset threshold and the type conversion relationship meets the preset logical continuity conditions, the cognitive state label of the silence interval is parsed as purposeful thinking.

[0041] In this embodiment, to accommodate the different operating rhythms of various students, the system uses a Gaussian Mixture Model (GMM) to establish an individualized time distribution baseline for each student. Specifically, the system removes extreme values, i.e., excessively long pauses above the 95th percentile and excessively short intervals below the 5th percentile. For the remaining regular interval samples, the Expectation-Maximization (EM) algorithm is used to fit the GMM. The number K of the mixture components is automatically determined using the Bayesian Information Criterion (BIC) to balance the model's fit and complexity. For each interval sample t, the probability that it belongs to the k-th Gaussian component is described by the following probability density function:

[0042] P(t|μ _K , σ _K )=(1 / (sqrt(2×π×σ _K ))×exp(-((t-μ _K ) 2 ) / (2×(σ _K ) 2 ));

[0043] Where, μ _K and σ _K Let be the mean and standard deviation of the k-th component, respectively, and π be the mathematical constant pi. Based on this model, the system calculates the current interval duration t. _i Relative to the primary reference component (usually the one whose mean is closest to t) _i Standardized deviation z of the components _interval Its calculation can be expressed as the following formula:

[0044] z _interval =(t _i -μ _ref ) / σ _ref ;

[0045] Or rather, μ _ref σ _ref These correspond to the mean and standard deviation of the main reference component, respectively.

[0046] When z _interval If the interval is less than a preset threshold, such as 1.5, it indicates that the interval is within the student's normal pace. At this point, the system further checks the logical continuation conditions.

[0047] Optionally, the system's computational logic tolerance L _logic If the preceding and following operations belong to a predefined pair of operations (e.g., setting parameters - running an experiment), L _logic The base score is set to 0.8; if subsequent parameter values ​​are numerically related to previous results, L... _logic Increase by 0.15. When L _logic Greater than 0.7 and z _interval When the value is less than 1.5, the system determines that the student is engaging in effective thinking related to the task and marks it as purposeful thinking.

[0048] Step 203: If a system loading flag or network latency record exists during the silence interval, the cognitive state label of the silence interval is parsed as system waiting;

[0049] If the standardization deviation is greater than a preset threshold, extract features of subsequent operations immediately adjacent to the silence interval;

[0050] When the subsequent operation is a rollback operation, a repetitive operation, or a request for help operation, the cognitive state label of the silence interval will be interpreted as confusion and stagnation.

[0051] When subsequent operations have no logical connection with preceding operations and the operation parameters exhibit random distribution characteristics, the cognitive state label of the silence interval is interpreted as distraction detachment.

[0052] In this embodiment, the system excludes non-human factors and prioritizes identifying the system's waiting state by matching loading events or network timeout records in the system log. For z _interval For pauses exceeding a preset threshold (e.g., 1.5, or even greater than 3.0), the system needs to distinguish between confusion and distraction. If subsequent actions involve undoing, checking help, or invalid repetitive attempts, it suggests the student encountered cognitive obstacles during the pause, thus classifying it as a confused pause. Conversely, if subsequent actions jump to irrelevant modules (logical break) and parameter settings deviate from a reasonable range (random distribution), it is determined that the student's attention has strayed from the current task, and this is marked as a distraction.

[0053] Furthermore, in order to handle cases with blurred boundaries, such as z _interval For values ​​between 1.5 and 2.0, or where both confusion and distraction characteristics are weakly present, this embodiment provides an optional processing method based on fuzzy logic. The system calculates the fuzzy membership degree of this interval to each cognitive state. For example, the membership degree μ for purposeful thinking. _think The calculation is as follows:

[0054] μ _think =max(0, 1-(z) _interval / 1.5))×L _logic ;

[0055] Membership degree μ belonging to confusion stagnation _confuse The calculation is as follows:

[0056] μ _confuse =min(1, z) _interval / 3.0)×C _confuse ×(1-D _distract );

[0057] Among them, C _confuse To clarify the strength of evidence, D _distract To assess the strength of distraction evidence, both are calculated based on a weighted score of subsequent operational features, with max(...) and min(...) corresponding to the maximum and minimum value functions, respectively. Based on this, the system determines the master state label using the normalized membership vector, preserving the uncertainty information in the state determination.

[0058] Step 204: Invoke the cognitive state label node. The cognitive state label node contains the parsed cognitive state label, the corresponding silence interval duration, and the membership vector of the silence interval to different cognitive states. Insert the corresponding cognitive state label node between each pair of adjacent operation events in the normalized operation event sequence. Perform unified symbol recoding on the operation events and cognitive state label nodes to generate a semantically enhanced operation sequence that alternately contains operation behavior information and cognitive semantic information.

[0059] Accordingly, the system encapsulates the parsing results into structured cognitive state label nodes. These nodes not only contain main labels, such as confusion and stagnation, but also retain the original physical quantities (duration) and ambiguity quantities (membership vectors), providing information input for subsequent algorithms.

[0060] Furthermore, the system physically inserts the node into the original sequence in operation E. _i With E _i+1 A unified encoding is used between them. For example, the operation type is encoded as O. _1 O _2 The state type is coded as S _1 S _2 The generated semantically enhanced operation sequence is formally represented as: O _A →S _Think →O _B →S _Confuse →O _C . Above, E _i With E _i+1 Represents adjacent operation nodes in the original normalized operation event sequence, where i is the index identifier of the operation node in the original sequence; O _1 O _2 O _A O _B O _C S represents different specific operation types in the sequence; _1 S _2 S _Think S _Confuse Indicates the state type encoding, for example, S _Think It can represent a state of thinking, S _Confuse It can indicate a state of confusion.

[0061] Based on this, the alternating structure allows subsequent sequence mining algorithms to simultaneously capture explicit operational patterns and implicit cognitive rhythms.

[0062] This solution employs a temporal-enhanced cognitive state analysis approach to address the problem of misjudgment of cognitive states caused by neglecting the semantics of operational intervals. By extracting operational context features and constructing an individualized temporal distribution baseline model, silence intervals are mapped to semantic labels such as purposeful thinking and confused stagnation, thus achieving an explicit expression of implicit cognitive processes.

[0063] On the other hand, this paper describes alternative technical solutions for identifying operational intentions and matching policy patterns, particularly how to quantitatively distinguish between students' scientific exploration and random trial intentions through multi-dimensional feature extraction and weighted fusion algorithms, transforming unstructured operation sequences into computable policy feature vectors. Specifically, this solution includes:

[0064] Step 301: Scan the semantically enhanced operation sequence and identify consecutively occurring subsequences of the same type of operation; when the length of the subsequence of the same type of operation exceeds the threshold and the operation parameters within the subsequence change numerically, mark the subsequence of the same type of operation as a repeated operation segment.

[0065] In this embodiment, the system employs a sliding window mechanism to traverse semantically enhanced operation sequences. To filter out accidental repetitive operations, the system sets a preset threshold, such as 3 times. Only when the window consecutively contains at least 3 operation events with the same operation type code (e.g., all setting heating temperature) but different parameter values ​​(e.g., 50°C, 60°C, and 70°C respectively), is the subsequence extracted as a repetitive operation segment. If the parameter values ​​are the same and the time interval is very short, it has already been identified as system jitter in the preprocessing stage and is not processed in this step.

[0066] Step 302: Extract multidimensional features from each repeated operation segment as exploration intent features. Exploration intent features include:

[0067] Regularity of parameter changes is used to characterize the monotonic trend and step size consistency of operational parameter changes within a segment;

[0068] Observe behavioral characteristics to characterize whether a silent interval marked as purposeful thinking occurs after each operation and the sufficiency of the observation duration;

[0069] The parameter space coverage feature is used to characterize the coverage range and distribution uniformity of the operating parameter values ​​over the reasonable range of parameter values.

[0070] The system quantifies the essence of scientific exploration from three dimensions. Correspondingly, for the regularity of parameter changes, the system calculates the monotonicity index of parameter changes. Specifically, assuming a segment contains n operations and the sequence of parameter changes is Δv, the monotonicity index is defined as the ratio of the length of the longest continuous segment where the signs of parameter changes remain consistent (all positive or all negative) to the total number of changes (n-1). Furthermore, the system calculates the step size consistency index C._step This indicator reflects whether the parameter adjustments are equidistant. Its calculation formula is:

[0071] C _step =1 / (1+(σ _step / μ _step ));

[0072] Where, σ _step μ is the standard deviation of the parameter variation step size. _step Let U be the mean of the parameter variation step size. Further, for the parameter spatial coverage characteristics, the system calculates the distribution uniformity U. _dist Specifically, the parameter value space is divided into m intervals, and the frequency p falling into each interval is counted. _j Calculate the normalized entropy value:

[0073] U _dist =(-Σ(p _j ×ln(p _j ))) / ln(m);

[0074] Above, U _dist The closer it is to 1, the more uniformly the parameter values ​​are distributed in space, which conforms to the characteristics of systematic traversal.

[0075] Based on this, for observed behavioral characteristics, the proportion R of silence intervals following all operations within a statistical segment that was analyzed as purposeful thinking was determined. _think The ratio of the average duration of thinking intervals to the baseline mean is used to determine whether students are actively thinking.

[0076] Step 303: Calculate the monotonicity index and step size consistency index of parameter changes in the repeated operation segment; based on the monotonicity index, step size consistency index, observation behavior characteristics and parameter space coverage characteristics, calculate the exploration intention index using a weighted summation method; when the exploration intention index is greater than the classification threshold, the operation intention type of the repeated operation segment is determined to be scientific exploration, otherwise it is determined to be random trial.

[0077] In this embodiment, the system fuses the above features using a linear weighted model to calculate the exploration intent index E. _index The calculation formula is as follows:

[0078] E _index =w _1 ×S _regular +w _2 ×S _observe +w _3 ×S _coverage ;

[0079] Among them, S _regular S is the weighted aggregate value of the monotonicity index and the step size consistency index._observe For R _think S is a weighted aggregate value of observation sufficiency. _coverage This is a weighted aggregate value based on coverage and distribution uniformity. The weighting coefficient is w. _1 w _2 w _3 The initial value can be set to 0.35, 0.40, or 0.25, corresponding to the dominant weight of observation behavior in determining exploration intent.

[0080] As an optional implementation, the classification threshold is not a fixed value, but is optimized and determined using the Receiver Operating Characteristic (ROC) curve method. Specifically, a logistic regression model is trained using a pre-labeled set of expert samples, and the threshold corresponding to the maximum point of the Youden index (the difference between the true positive rate and the false positive rate) is selected as the classification threshold, maximizing the discriminative power while ensuring recognition accuracy.

[0081] Step 304: Match the semantically enhanced operation sequence with the attached operation intent type with multiple typical policy patterns in the typical policy pattern library one by one;

[0082] For each typical strategy pattern, calculate the inclusion degree, operation sequence consistency, and cognitive rhythm similarity between the student sequence and the subsequence of that typical strategy pattern;

[0083] Matching scores are generated based on subsequence inclusion, operational order consistency, and cognitive rhythm similarity. The matching scores for each typical strategy pattern are then combined into a strategy feature vector.

[0084] Alternatively, compare the complete operational process with a clear operational purpose with the standard strategy templates in the typical strategy pattern library one by one.

[0085] For each standard template, check whether the student's operation contains the template's steps, whether the order of steps is consistent with the template, and whether the operation rhythm (e.g., how long it takes to complete a step, when to make adjustments) is similar to the template.

[0086] Based on the above three points, a matching score is calculated, and the scores of all templates are combined to form feature data.

[0087] Correspondingly, the system introduces a typical strategy pattern library, transforming complex student behavior sequences into computer-processable mathematical vectors. The system employs the Longest Common Subsequence (LCS) algorithm from dynamic programming to calculate the subsequence inclusion degree between the student sequence and the pattern sequence, reflecting structural similarity. Furthermore, the system calculates the cognitive rhythm similarity (Rhythm). _sim Specifically, the cosine similarity formula is used, namely:

[0088] Rhythm _sim =(V _student ×V_pattern ) / (|V _student |×|V _pattern |);

[0089] Among them, V _student This represents the distribution vector of students' cognitive states within the matched segment, such as [thinking percentage, confusion percentage, ...], or [thinking percentage, confusion percentage, ..., completion percentage]. The completion percentage represents the proportion of time spent organizing and submitting after completing the task, such as saving results and submitting assignments; V _pattern This is the standard cognitive distribution vector corresponding to this typical pattern. Based on this, the system arranges the comprehensive matching scores for N patterns in the library into an N-dimensional vector, which is the policy feature vector.

[0090] Furthermore, this solution constructs a quantitative model of exploration intent, addressing the problem that traditional sequence mining cannot distinguish between scientific exploration and random trial and error. By calculating characteristics such as the monotonicity of parameter changes, step size consistency, and parameter space coverage, systematic inquiry behavior is defined mathematically, solving the problem that cognitive depth cannot be identified solely based on operational symbols.

[0091] Some embodiments provide specific implementation methods for quantifying and evaluating decision hesitation, such as how to quantify cognitive hesitation in the decision-making process through micro-behavioral analysis and establish a multi-dimensional strategy effectiveness evaluation model. Specifically, this method includes:

[0092] Step 401: Scan the semantically enhanced operation sequence to locate the key decision-making branch points where students face multiple path choices;

[0093] For each key decision-making branch, a decision window is set, and multi-dimensional micro-behavioral features of the decision are extracted within the decision window;

[0094] Decision-making micro-behavioral characteristics include: the number of times the alternatives are previewed before making a final choice, the number of times the operation is canceled or rolled back after it is initiated, the frequency of switching between different options, and the duration of the pause before making a final choice.

[0095] Correspondingly, key decision branches are operational nodes in the experimental process where there are multiple valid branches or parameter selection spaces. The system can match nodes based on a predefined task knowledge graph, or it can use a data-driven approach to identify nodes in the operation sequence where the divergence of subsequent paths exceeds a threshold as potential decision points, adapting to open-ended exploration tasks. After locating a decision point, the system extracts several operations before and after that point, for example, three operations before and after, to form a decision window, and statistically analyzes micro-behavioral data within the window.

[0096] Step 402: Standardize the micro-behavioral characteristics of the decision-making process;

[0097] The decision hesitation index is calculated based on standardized micro-behavioral characteristics of decision-making. The number of cancellations or rollbacks, the switching frequency, and the pause duration after cognitive state correction contribute positively to the decision hesitation index, while the number of previews contributes negatively to the decision hesitation index.

[0098] In other words, the subtle behavioral characteristics observed in students' decision-making are uniformly standardized, and the degree of decision hesitation is calculated based on this.

[0099] Among them, the number of times students cancel / rollback operations and the frequency of behavior switching, combined with the pause time after cognitive state correction, increase the decision hesitation index; conversely, the more times previewing in advance, the lower the hesitation.

[0100] Furthermore, the system performs Z-score standardization on the feature values ​​of each micro-behavior to eliminate dimensional differences. Further, the system calculates the decision hesitation index H. _decision Its design follows psychological principles, namely repeated cancellation (N... _cancel ), hovering switch (F _switch ) and prolonged pauses (T _pause ) indicates hesitation, while multiple previews (P) _preview This indicates a prepared decision. The calculation formula is as follows:

[0101] H _decision =α×norm(N _cancel )+β×norm(F _switch )+γ×κ _state ×norm(T _pause )-δ×norm(P _preview );

[0102] Where norm represents the standardization function; α, β, γ, and δ are weighting coefficients, determined through correlation analysis with students' post-hoc subjective confusion reports, for example, they can be set to 0.30, 0.25, 0.30, and 0.15 respectively, κ _state Corresponding correction factor. This formula introduces a negative term, namely the number of previews, to distinguish between blind hesitation and cautious decision-making.

[0103] Step 403: Retrieve the cognitive state label corresponding to the pause duration before making the final choice in the semantically enhanced operation sequence;

[0104] If the cognitive state label is "purposeful thinking", then reduce the contribution weight of this pause duration in calculating the decision hesitation index.

[0105] If the cognitive state label is confusion and stagnation, then the contribution weight of this pause duration in calculating the decision hesitation index is increased.

[0106] The system introduces a correction coefficient κ. _state The γ term in the formula is dynamically adjusted. If a pause is interpreted as purposeful thinking, it indicates that the student is engaging in effective reasoning; in this case, κ is set. _state =0.3, reducing its contribution to hesitation; if the pause is interpreted as a confused stagnation, it indicates that the student is encountering an obstacle, so κ is set. _state =1.2, amplify its contribution; if it is a distraction detachment, set κ. _state =0.5. The semantic-based correction mechanism allows the hesitation index to reflect the degree of cognitive difficulty, rather than the time spent.

[0107] Step 404: Generate policy effectiveness evaluation results by combining policy feature vectors, including a comprehensive evaluation from the following three dimensions:

[0108] Strategy type fit is calculated based on the degree of matching between the strategy feature vector and the typical strategy patterns of high-performing groups;

[0109] The smoothness of strategy execution is calculated based on the number and distribution density of high hesitation level decision points in the semantically enhanced operation sequence, where high hesitation level decision points are decision points whose decision hesitation index exceeds the threshold.

[0110] The effectiveness of the strategy output is calculated based on the correctness of the conclusions of the experimental task and the efficiency of the completion time.

[0111] Furthermore, the system's overall performance score E is calculated. _strategy Among them, the strategy type fit S _match The strategy execution smoothness F is calculated using weighted cosine similarity, with weights determined by the average performance of each typical pattern. _fluency Based on the proportion of highly hesitant decision points, the calculation formula can be described as follows:

[0112] F _fluency =1-(N _high_hesitation / N _total_decision );

[0113] Where, N _high_hesitation The corresponding number of highly hesitant decision points, N _total_decision Total number of corresponding decision points. Strategy output effectiveness O. _outcome It combines the accuracy of conclusions with time efficiency. Performance evaluation result E _strategy The weighted summation yields the following formula:

[0114] E _strategy =λ _1 ×S _match +λ _2 ×F _fluency +λ _3 ×O _outcome ;

[0115] This assessment model not only focuses on the results (effectiveness) but also on the quality of the process (adaptability and fluency), providing multi-dimensional diagnostic basis for subsequent personalized guidance.

[0116] In other embodiments, optional implementation schemes for cognitive load perception and progressive guidance are described, particularly how to utilize multimodal features to perceive students' cognitive load levels in real time, and thereby generate dynamic, progressive strategy transition paths to address the problem of cognitive overload caused by rote learning. Correspondingly, this scheme can be implemented using the following steps:

[0117] Step 501: Extract real-time load features from the semantically enhanced operation sequence using a sliding window mechanism. The real-time load features include:

[0118] Operation frequency characteristics, error rate characteristics, silent pattern characteristics based on cognitive state label statistics, and operation complexity characteristics;

[0119] Based on real-time load characteristics, the real-time cognitive load level is calculated using a pre-built cognitive load estimation model.

[0120] Based on this, the system monitors students' cognitive levels in real time. The system can set a sliding time window, such as 60 seconds, and calculate the following four-dimensional features within the window:

[0121] The frequency of operation reflects the urgency of the action;

[0122] Error rate characteristics statistics show the percentage of errors or invalid operations reported by the system;

[0123] The silent mode feature is based on the analysis results, which statistically represents the proportion of time spent in a state of confusion and stagnation.

[0124] Operational complexity features are based on the inherent difficulty weights of the current task steps.

[0125] Optionally, fuzzy inference can be used to fuse heterogeneous features into a unified load level. The system defines a fuzzy rule base; for example, if the error rate membership degree is high, the confusion / stagnation ratio membership degree is high, and the operation frequency membership degree is low, then the cognitive load level is high. Through defuzzification calculation using fuzzy inference, continuous load estimates (0 to 1) and discrete levels (low, moderate, high) are output. This mechanism ensures that the guidance system does not interfere with students when they are busy or deep in thought.

[0126] Step 502: Select the target strategy from the typical strategy pattern library that has higher performance than the current strategy and whose similarity to the current strategy meets the preset conditions.

[0127] By comparing the current strategy with the target strategy, the differences are broken down into several strategic elements that can be improved independently.

[0128] Based on the difficulty of improving the strategy elements and the real-time cognitive load level, the strategy elements are assigned to different transition stages, generating a transition path containing multiple progressive stages.

[0129] The system follows the zone of proximal development (ZPD) theory for path planning. When selecting a target strategy, the system does not directly choose the highest-performing pattern from the database; instead, it calculates the Euclidean distance between candidate patterns and the student's current strategy feature vector. The system then selects strategies whose distance is less than a preset upper limit and whose performance improvement is greater than ΔP. _min The model sorts targets by performance-distance composite score and selects the optimal, achievable strategy as the goal. In other words, ΔP... _min This indicates the threshold for performance improvement.

[0130] Furthermore, when generating transition paths, the system performs strategy difference decomposition. For example, the differences are broken down into independent strategy elements such as adjusting the operation sequence, optimizing parameter settings, and cultivating observation habits. For each element, the system assesses its improvement difficulty in conjunction with key decision-making weaknesses. To prevent cognitive overload, the system sets a single-stage improvement capacity limit, such as two items.

[0131] The path planning algorithm assigns less difficult elements to the first stage and more difficult elements to subsequent stages. For example, the transition path can be planned as: Stage 1 (optimize heating parameter settings) → Consolidation period → Stage 2 (adjust the frequency of observation results) → Target strategy. The system will only unlock the guidance content for the next stage after the student's performance in the current stage consistently meets the preset standards.

[0132] Building upon this foundation, this application establishes a decision hesitation index and a progressive path planning mechanism based on micro-behavioral analysis. By integrating micro-behavioral features such as preview and cancellation to quantify the degree of decision-making difficulty, and dynamically adjusting the difficulty and timing of guidance in conjunction with real-time cognitive load levels, it solves the problems of traditional guidance lacking specificity and easily leading to cognitive overload in students.

[0133] In other embodiments, alternative implementations for offline construction of a typical strategy pattern library are described. The offline, knowledge discovery process based on historical group data provides a reference standard for online evaluation. Accordingly, this embodiment includes:

[0134] Step 601: Filter historical operation sequence data of high-performing student groups;

[0135] Construct an operation semantic similarity matrix that includes operation type similarity and operation object association.

[0136] The system calculates each student's overall performance score P based on historical experiment records. _perform(Integrating accuracy, time efficiency, and operational redundancy), the top 25% of students are selected as the high-performing group. The system constructs an operational semantic similarity matrix to address the problem that traditional mining algorithms can only match identical symbols. For any two operational symbols o... _i and o _j Its semantic similarity Sim(o _i o _j The calculation formula is as follows:

[0137] Sim(o _i o _j )=w _type ×S _type +w _obj ×S _object +w _ctx ×S _context ;

[0138] Among them, S _type For operation type similarity, 1.0 is given for operations within the same major category and subcategory, and 0.7 is given for operations within the same major category but different subcategories; S _object Functional affinity of the manipulated objects (based on a predefined object graph); S _context For contextual co-occurrence similarity, it is calculated as the frequency of two operations occurring adjacently in the historical sequence, with a weighting coefficient w. _type w _obj w _ctx These can be set to 0.4, 0.3, and 0.3 respectively. Sim(...) represents the similarity calculation function. This formula treats setting temperature and setting pressure as similar operation nodes to some extent, improving the generalization ability of pattern mining.

[0139] Step 602: Based on the operation semantic similarity matrix, a sequence pattern mining algorithm with a semantic similarity merging mechanism is used to mine high-frequency and efficient operation skeleton patterns from historical operation sequence data.

[0140] Alternatively, a sequence pattern mining algorithm that combines the operation semantic similarity matrix with a semantic similarity merging mechanism can mine operation skeleton patterns from historical operation sequence data and select high-frequency and efficient pattern results.

[0141] Optionally, the system employs an improved PrefixSpan algorithm. During pattern growth, when determining candidate operations... _candidate When determining whether a prefix can be used as a successor, the system checks not only an exact match but also Sim(o) _candidate o _existing Is it greater than a similarity threshold, for example, 0.75? _existing If a similar operation is found, it is merged into an existing pattern branch.

[0142] Furthermore, the system introduces performance-weighted support, meaning that the support calculation for a pattern no longer simply counts the number of occurrences, but rather sums the performance scores of students who exhibit that pattern. This ensures that the discovered operational skeleton patterns are not only high-frequency, but also genuinely represent the key behavioral paths leading to high performance.

[0143] Step 603: Extract specific instance fragments of the operation skeleton pattern in the historical operation sequence data;

[0144] Analyze the distribution patterns and transition characteristics of cognitive state labels in example fragments, identify typical cognitive rhythm features associated with operational skeleton patterns, and store typical cognitive rhythm features in a typical strategy pattern library.

[0145] In this embodiment, a typical strategy includes not only what to do (operational skeleton) but also how to think (cognitive rhythm). The system reviews historical data, extracts specific instance fragments corresponding to each skeleton pattern, and statistically labels the cognitive state within each fragment. For example, the system calculates a transition probability matrix where purposeful thinking and operation alternate. If instances of a certain pattern generally exhibit a strong rhythm of operation → thinking → operation, then this feature is stored in the database as a typical cognitive rhythm feature. Based on this, the constructed pattern library includes: pattern ID, operation sequence template, typical cognitive rhythm, average performance level, and applicable task type.

[0146] One example provides auxiliary computational models and alternative implementation methods, particularly screening criteria for high-performing groups, time-efficiency normalized computation methods, and data-driven methods for identifying key decision points. These include:

[0147] Step 801: When screening historical operation sequence data of high-performing student groups, a multi-dimensional comprehensive performance score calculation model is adopted.

[0148] The system needs to quantify students' overall performance. Specifically, the system collects the accuracy rate (A) of students' experimental conclusions. _correct (i.e., the degree of matching between the conclusion and the standard answer), task completion time efficiency E _time (The reciprocal of the completion time is standardized) and operational redundancy R _redundant (The proportion of valid operations to total operations). The system uses a weighted summation formula to calculate the overall performance score P. _perform :

[0149] P _perform =w _p1 ×A _correct +w _p2 ×E _time +w _p3 ×R _redundant ;

[0150] Wherein, the weighting coefficient w _p1 w _p2 w _p3 These values ​​can be set to 0.5, 0.3, and 0.2 respectively to reflect the dominant role of the result's correctness. The system is based on P. _perform Students were ranked, and the top 25% with a score of at least 0.6 were selected as the sample source.

[0151] Step 802: When calculating the effectiveness of the strategy output, a time efficiency conversion model based on population distribution is adopted.

[0152] In this embodiment, for completion time efficiency, the system does not directly use absolute time, but converts it into a relative efficiency index T. _efficiency The system statistically analyzes the completion time distribution of the same task group and extracts the 5th percentile T. _min and the 95th percentile T _max T _efficiency The calculation formula is as follows:

[0153] T _efficiency =1-(T _complete -T _min ) / (T _max -T _min );

[0154] Among them, T _complete This represents the actual completion time for the current student. If the calculated result is greater than 1, it is truncated to 1; if it is less than 0, it is truncated to 0. This formula helps eliminate differences in the order of magnitude of time between different experimental tasks, making the time indicators comparable across different scenarios.

[0155] Step 803: As an alternative to locating decision points using a predefined knowledge graph, a data-driven method based on information entropy is used to locate key decision branch points.

[0156] Optionally, for more exploratory open tasks (i.e., localization steps), the system provides an implementation method that does not rely on expert knowledge. Accordingly, the system scans massive historical operation sequences, targeting each operation node. _i Calculate the distribution of the types of the next operations. Assume the probability distribution of the subsequent operation types is p. _1 p _2 , ..., p _K The system calculates the entropy H of the subsequent operation of this node. _next ,Right now:

[0157] H _next =-Σ(p _K ×ln(p _K ));

[0158] When H _nextWhen the divergence threshold exceeds a preset threshold (e.g., 1.5), it indicates that the behavioral paths of different students have diverged after that node, and that node is marked as a critical decision branch. This scheme can automatically identify operational steps that experts failed to foresee but where cognitive divergence actually exists.

[0159] According to one aspect of this application, some methods of the present invention may also be:

[0160] Optionally, obtain the time interval sequence with contextual features for each student, and calculate the 95th percentile P of all time interval values ​​for that student. _95 With the 5th percentile P _5 Intervals exceeding P _95 Records marked as extremely long pause samples are considered unconventional and may correspond to task interruptions, extended periods of absence, or other abnormal situations, making them unsuitable for inclusion in regular operation rhythm modeling. Intervals shorter than P are also considered. _5 Records with response times close to the system's minimum (typically below 100 milliseconds) are marked as ultra-short interval samples. These samples may be automatically triggered continuously by the system rather than being actively operated by the student. Ultra-long pause samples and ultra-short interval samples are removed from the modeling data and saved separately in an extreme interval sample set for subsequent special state analysis. The remaining samples constitute the regular interval modeling dataset, serving as input for fitting the Gaussian mixture model.

[0161] Optionally, the number of mixture components is adaptively selected for the interval duration data in the conventional interval modeling dataset. Candidate Gaussian mixture models are constructed with three configurations: K equal to 2, 3, and 4 components. For each configuration, the expectation-maximization algorithm is performed to estimate the parameters. For each candidate model, the Bayesian Information Criterion (BIC) value is calculated. _K The calculation formula is:

[0162] BIC _K =-2×L _K +m _K ×ln(n);

[0163] In the formula, L _K m represents the log-likelihood of the model when the number of components is K. _K The number of free parameters in the model is 3K-1 (for a Gaussian mixture model with K components, the number of free parameters is 3K-1, including K means, K variances and K-1 mixture weights), and n represents the number of samples in the regular interval modeling dataset.

[0164] Accordingly, the number of components with the smallest BIC value is selected as the optimal configuration K. _opt This criterion strikes a balance between model goodness of fit and complexity, avoiding overfitting. It outputs the optimal number of components K. _opt This information is then passed to the next step for formal parameter estimation.

[0165] Optionally, the dataset is modeled based on the regular interval and the optimal number of components K. _opt The complete Gaussian mixture model parameter estimation process is performed. The K-means++ algorithm is used to perform K-means estimation on the interval time data. _opt Clustering, using the cluster centers as the mean μ of each Gaussian component. _K The initial value is the standard deviation of each type of sample, which is used as the component standard deviation σ. _K The initial value is set using the proportion of each type of sample as the mixing weight π. _K The initial values, where k ranges from 1 to K. _opt The above refers to the K-means++ algorithm, which is an optimization algorithm for K-means clustering.

[0166] Furthermore, the expectation step and the maximization step are performed alternately. In the expectation step, for each interval sample t... _i Calculate the posterior probability γ of its belonging to each component. _ik The calculation formula is:

[0167] γ _ik =(π _K ×N(t _i |μ _K , σ _K )) / Σ _j (π _j ×N(t _i |μ _j , σ _j ));

[0168] Wherein, N(t) _i |μ _K , σ _K ) indicates that μ _K For the mean, σ _K A Gaussian distribution with standard deviation at t _i The probability density value at point N(t) _i |μ _j , σ _j Similarly, π _K π _j For the corresponding mixed weights, summation symbol Σ _j This indicates that for all components j from 1 to K _opt Summation. In the maximization step, the parameters of each component are updated based on the posterior probability, which can be expressed by the following formula:

[0169] Updated mean μ _K_new =Σ _i (γ _ik ×t _i ) / Σ _i (γ _ik);

[0170] Updated standard deviation σ _K_new =sqrt(Σ _i (γ _ik ×(t _i -μ _K_new ) 2 ) / Σ _i (γ _ik ));

[0171] Updated mixed weights π _K_new =Σ _i (γ _ik ) / n;

[0172] The iteration termination condition is set as follows: the change in log-likelihood is less than a threshold ε (which can be 0.001) or the number of iterations reaches an upper limit (e.g., 100 iterations). Upon convergence, the final parameter estimates for each component are output, including the mean set {μ}. _1 μ _2 , ..., μ _K_opt}, Standard deviation set {σ _1 , σ _2 , ..., σ _K_opt} and the mixed weight set {π _1 , π _2 , ..., π _K_opt}, the estimated results of the constituent parameters. Where, the subscripts... _K_opt The corresponding optimal number of components.

[0173] Optionally, semantic annotation and boundary delineation are performed on each Gaussian component in the component parameter estimation results. Based on the mean μ... _K The components are sorted from smallest to largest. Usually, the smallest mean component corresponds to the fast and continuous operation mode (students are proficient in performing coherent operations), the medium mean component corresponds to the normal thinking interval mode (students perform routine cognitive processing), and the largest mean component corresponds to the longer pause mode (students perform deep thinking or encounter mild confusion).

[0174] Calculate the boundary threshold θ between adjacent components _boundary Using the equal-probability crossover method, for adjacent components k and k+1, the interval length that makes the probability densities of the two components equal is calculated as the boundary point:

[0175] π _K ×N(θ|μ _K , σ _K )=π _k+1 ×N(θ|μ _k+1 , σ _k+1 );

[0176] The boundary threshold θ is obtained through numerical solution._k，k+1 For K _opt Components, total K calculated _opt -1 boundary thresholds constitute the boundary threshold sequence {θ} _1，2 θ _2，3 , ...}. Or, the boundary threshold sequence is {θ _1，2 θ _2，3 , ..., θ _K_opt-1，K_opt}

[0177] Where θ represents the time interval boundary between adjacent Gaussian components, N(θ|μ _K , σ _K ) indicates that μ _K For the mean, σ _K The probability density value of a Gaussian distribution with standard deviation π at θ is given by π. _k+1 N(θ|μ) represents the mixing weight of the (k+1)th Gaussian component. _k+1 , σ _k+1 ) indicates that μ _k+1 For the mean, σ _k+1 The probability density value of a Gaussian distribution with standard deviation at θ is given.

[0178] By integrating the number of components, parameters of each component (mean, standard deviation, and mixture weight), semantic labels of components, and boundary threshold sequences, an individual time baseline model parameter set is output. This parameter set characterizes the personalized operational time rhythm features of each student, providing an individualized reference standard for subsequent semantic classification of silence intervals.

[0179] Furthermore, some methods of the present invention may also be as follows:

[0180] Iterate through each interval record in the time interval sequence with contextual features, and perform the identification and judgment of the system waiting state. This state is caused by system factors rather than student cognitive activities and should be excluded first.

[0181] Check for system-level waiting flags within the interval, including: page load event records (identified by the system response status field in the operation log), network request timeout records, experimental resource loading flags, and system prompt pop-up display periods. If the interval overlaps with any system event, and the overlap duration exceeds 50% of the total interval duration, the interval is marked as a system waiting state with a confidence level of 1.0. After marking, this interval will no longer participate in subsequent cognitive state classification.

[0182] For interval records that are not marked as system wait, their complete feature information is retained to form a set of intervals to be classified, which will then enter the subsequent cognitive state determination process.

[0183] Furthermore, for each interval record in the set of intervals to be classified, the standardized deviation of the interval duration from the individual's regular distribution is calculated by combining the corresponding student's individual time baseline model parameter set.

[0184] Based on the interval duration t _i Determine the baseline component to which it most likely belongs. For each component k in the individual time baseline model, calculate the interval duration t. _i The posterior probability p of belonging to this component _K The component with the highest posterior probability is taken as the reference component k. _ref Extract the mean μ of this component. _ref With standard deviation σ _ref .

[0185] Accordingly, the standardized deviation z is calculated. _i Its formula is:

[0186] z _i =(t _i -μ _ref ) / σ _ref ;

[0187] This deviation reflects the degree of abnormality of the current interval duration relative to the student's personal routine. Positive values ​​indicate that the interval is longer than the individual's routine, negative values ​​indicate that the interval is shorter than the individual's routine, and the larger the absolute value, the greater the degree of deviation.

[0188] Simultaneously calculate the overall deviation z _global The population mean μ of all regular intervals of the individuals _global Compared with the population standard deviation σ _global For reference, the specific formula is as follows:

[0189] z _global =(t _i -μ _global ) / σ _global ;

[0190] Based on this, z _i With z _global They are all attached to the interval records, and the output is a set of intervals to be classified with deviation.

[0191] Furthermore, for the standardized deviation z in the set of unclassified intervals with deviation, _i Within the normal range (the judgment interval is set to z) _i For intervals less than 1.5, further determine whether they belong to a state of purposeful thinking.

[0192] Extract contextual features from interval records, focusing on analyzing the logical relationship between preceding and subsequent operations. The rules for determining logical relationship include: whether the subsequent operation is a natural continuation of the preceding operation (e.g., parameter fine-tuning, result viewing, execution of the next step); whether the subsequent operation belongs to the same task sub-objective category as the preceding operation; and whether the parameter settings of the subsequent operation reflect a response to the result of the preceding operation, such as adjusting variable values ​​based on observation results.

[0193] Constructing the logical continuity index L _logic A comprehensive evaluation of the above-mentioned correlation characteristics is conducted. If the preceding and following operations belong to a predefined pair of successive operations, such as setting parameters - running the experiment, observing results - recording data, L _logic The base score is set to 0.8; if the value of the subsequent operation parameter is numerically related to the result of the previous operation, such as a transformation in which the value of the subsequent parameter equals the value of the previous result, L _logic Increase by 0.15; if there are no operation cancellations or rollbacks recorded during the interval between operations, L _logic Increase by 0.05.

[0194] When z _i Less than 1.5 and L _logic When the value is greater than 0.7, the interval is marked as a state of purposeful thinking, indicating that the student engaged in effective cognitive processing related to the task during the pause. The status label and L... _logic Values ​​are appended to interval records, marked intervals are removed from the set to be classified, and the remaining intervals form the long interval set to be classified.

[0195] Furthermore, regarding the standardized deviation z in long-interval sets to be classified... _i Intervals greater than or equal to 1.5 are used to distinguish between two cognitive states: confusion stagnation and distraction detachment. Behavioral features of the first action (i.e., subsequent actions) after the interval are extracted to construct a set of confusion indicator features, as follows:

[0196] Feature F1 corresponds to the rollback operation indicator. It checks whether the subsequent operation is a rollback operation such as undo, return to the previous step, or reset. If so, F1 is set to 1; otherwise, it is 0.

[0197] Feature F2 corresponds to the repeat operation indicator. It checks whether the subsequent operation is of the same type as the previous operation but the parameter values ​​change very little or are the same. If so, F2 is 1; otherwise, it is 0.

[0198] Feature F3 corresponds to the help operation instruction. It checks whether the subsequent operation is a help-related operation such as viewing help, opening prompts, or consulting reference materials. If so, F3 is set to 1; otherwise, it is 0.

[0199] Feature F4 corresponds to the error operation indicator. Check whether subsequent operations trigger system error messages or are judged as invalid operations by the system. If so, F4 is set to 1; otherwise, it is 0.

[0200] Based on this, the strength of confusing evidence C is calculated. _confuse The formula is:

[0201] C _confuse =0.3×F1+0.25×F2+0.25×F3+0.2×F4; 0.3, 0.25, 0.25, and 0.2 are the corresponding weighting coefficients.

[0202] Accordingly, the set of distraction indicator features is extracted, namely:

[0203] Feature D1 corresponds to the logic break indicator. It checks whether the subsequent operation and the preceding operation belong to unrelated task modules. If so, D1 is 1; otherwise, it is 0.

[0204] Feature D2 corresponds to the random parameter indicator. It checks whether the parameter values ​​of subsequent operations exhibit random characteristics, that is, the deviation of the parameter values ​​from the center of the reasonable value range is greater than 2 standard deviations. If so, D2 is set to 1; otherwise, it is 0.

[0205] Feature D3 corresponds to the fast continuous indicator. It checks whether a subsequent operation is followed by multiple fast continuous operations and whether there is no logical connection between the operations. If so, D3 is 1; otherwise, it is 0.

[0206] Based on this, the strength of the evidence of distraction, D, is calculated. _distract The formula can be described as:

[0207] D _distract =0.4×D1+0.35×D2+0.25×D3; 0.4, 0.35, and 0.25 correspond to the available weight coefficients.

[0208] Based on this, compare C _confuse With D _distract The size of C _confuse Greater than D _distract And C _confuse If D is greater than 0.3, mark the interval as a confused stagnation state; if D _distract Greater than C _confuse And D _distract If the deviation is greater than 0.3, the interval is marked as a distraction-free state; if neither is greater than 0.3, the deviation is determined according to the standardized deviation z. _i Further determination, if z _i If the value is greater than 3.0, it tends to be marked as distracted and detached; otherwise, it is marked as confused and stagnant. Output the preliminary state marking results.

[0209] Furthermore, for the interval records where the boundaries are ambiguous in the preliminary state marking results, i.e., C _confuse With D _distract The difference is less than 0.15, or z _i Within the boundary range of 1.5 to 2.0, fuzzy membership degree calculation is performed, preserving the uncertainty information of state determination.

[0210] Additionally, a four-state fuzzy membership degree calculation rule is designed. For the purposeful thinking state, the membership degree μ... _think The calculation can be:

[0211] μ _think =max(0, 1-z) _i / 1.5)×L _logic ;

[0212] For a confused stagnation state, the membership degree μ _confuse The calculation is as follows:

[0213] μ _confuse =min(1, z) _i / 3.0)×C _confuse ×(1-D _distract );

[0214] For the distracted detached state, the membership degree μ _distract The calculation is as follows:

[0215] μ _distract =min(1, z) _i / 3.0)×D _distract ×(1-C _confuse );

[0216] For the system waiting state, the membership degree μ _system It has been determined that this value is 0 for non-system wait intervals. Alternatively, it is 1 when the system is in a wait state, and 0 otherwise.

[0217] The membership vector is normalized so that the sum of the membership degrees of the four states is 1, as follows:

[0218] μ _K_norm =μ _K / (μ _think +μ _confuse +μ _distract +μ _system );

[0219] The state with the highest normalized membership degree is taken as the primary state label, and the state with the second highest membership degree is taken as the secondary state label. The primary state label, secondary state label, and membership vector are appended to each interval record. All intervals (including the labeled system waiting intervals) are integrated to output a time interval sequence with cognitive state labels. Each interval contains the primary state, secondary state, membership vector, original interval duration, and contextual features.

[0220] The optional embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solution of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for evaluating and guiding learning strategies based on experimental operation sequence pattern mining, characterized in that, include: Collect raw operation log data generated by students on the experimental platform and reconstruct it into a standardized sequence of operation events based on the operation timestamps; Extract the silence interval features between adjacent operations in the normalized operation event sequence, combine them with operation context features to parse the cognitive state labels corresponding to the silence intervals, and embed them into the normalized operation event sequence to generate semantically enhanced operation sequences. Identify repetitive operation segments in semantically enhanced operation sequences, calculate exploration intention features to distinguish operation intention types, and extract students' strategy feature vectors by combining typical strategy pattern libraries; Locate key decision branch points in semantically enhanced operation sequences, extract decision micro-behavioral features during the decision-making process, calculate decision hesitation index, and combine with policy feature vectors to generate policy effectiveness evaluation results; Real-time cognitive load levels are calculated based on semantically enhanced operation sequences. Combined with policy effectiveness evaluation results, a gradual transition path from the current policy to the target policy is planned, and personalized policy guidance suggestions are output. The process of extracting the silence interval features between adjacent operations in a normalized operation event sequence includes: calculating the timestamp difference between adjacent operation events in the normalized operation event sequence as the silence interval duration; extracting the operation context features corresponding to each silence interval, which include: the complexity level of the preceding operation, the type conversion relationship between the subsequent operation and the preceding operation, and the relative position ratio of the current silence interval in the sequence; wherein, the type conversion relationship includes continuation of the same type, cross-type switching and reverse rollback, and the silence interval feature is a composite feature that integrates the silence interval duration and the operation context features; The process involves embedding cognitive state labels into a normalized sequence of operational events to generate a semantically enhanced sequence of operations. This includes: calling a cognitive state label node, which contains the parsed cognitive state label, the corresponding silence interval duration, and the membership vector of the silence interval to different cognitive states; inserting a corresponding cognitive state label node between each pair of adjacent operational events in the normalized sequence of operational events; and performing unified symbol recoding on the operational events and the cognitive state label nodes to generate a semantically enhanced sequence of operations that alternately contains operational behavior information and cognitive semantic information. The process of identifying repetitive operation segments in a semantically enhanced operation sequence includes: scanning the semantically enhanced operation sequence and identifying consecutively occurring subsequences of the same type of operation; when the length of a subsequence of the same type of operation exceeds a threshold and the operation parameters within the subsequence change numerically, the subsequence of the same type of operation is marked as a repetitive operation segment. The calculation of exploration intent features distinguishes the types of operational intent. This includes extracting multi-dimensional features of each repetitive operation segment as exploration intent features. These features include: parameter change regularity features, which characterize the monotonic trend and step size consistency of operational parameter changes within the segment; observation behavior features, which characterize whether a silent interval marked as purposeful thinking occurs after each operation and the sufficiency of the observation duration; and parameter space coverage features, which characterize the coverage and distribution uniformity of operational parameter values ​​within the reasonable range of parameter values. Specifically, the calculation of exploration intent features distinguishes the types of operational intent, including: calculating the monotonicity index and step size consistency index of parameter changes in repeated operation segments; calculating the exploration intent index using a weighted summation method based on the monotonicity index, step size consistency index, observation behavior features, and parameter space coverage features; when the exploration intent index is greater than the classification threshold, the operational intent type of the repeated operation segment is determined to be scientific exploration, otherwise it is determined to be random trial.

2. The method according to claim 1, characterized in that, Extract student policy feature vectors by combining a typical policy pattern library, including: The semantically enhanced operation sequence with attached operation intent type is matched with multiple typical policy patterns in the typical policy pattern library respectively; For each typical strategy pattern, calculate the inclusion degree, operation sequence consistency, and cognitive rhythm similarity between the student sequence and the subsequence of that typical strategy pattern; Matching scores are generated based on subsequence inclusion, operational order consistency, and cognitive rhythm similarity. The matching scores for each typical strategy pattern are then combined into a strategy feature vector.

3. The method according to claim 1, characterized in that, Extracting micro-behavioral features of the decision-making process, including: Scan semantically enhanced operation sequences to locate key decision branches where students face multiple path choices; For each key decision-making branch, a decision window is set, and multi-dimensional micro-behavioral features of the decision are extracted within the decision window; Decision-making micro-behavioral characteristics include: the number of times the alternatives are previewed before making a final choice, the number of times the operation is canceled or rolled back after it is initiated, the frequency of switching between different options, and the duration of the pause before making a final choice.

4. The method according to claim 3, characterized in that, The calculation of the decision hesitation index specifically includes: Standardize the micro-behavioral characteristics of decision-making; The decision hesitation index is calculated based on standardized decision micro-behavioral characteristics. The number of cancellations or rollbacks, the switching frequency, and the pause duration after cognitive state correction contribute positively to the decision hesitation index, while the number of previews contributes negatively to the decision hesitation index.

5. The method according to claim 4, characterized in that, The pause duration, adjusted based on cognitive state, includes: The cognitive state label in the semantically enhanced operation sequence corresponding to the pause duration before making a final choice in the retrieval process; If the cognitive state label is "purposeful thinking", then reduce the contribution weight of this pause duration in calculating the decision hesitation index. If the cognitive state label is confusion and stagnation, then the contribution weight of this pause duration in calculating the decision hesitation index is increased.

Images