Learning behavior log segmentation method based on behavior evolution

Abstract

The invention provides a learning behavior log segmentation method based on behavior evolution, and belongs to the field of learning behavior analysis and data processing. Firstly, data preprocessing and feature extraction are carried out, weight coefficients are allocated to extracted features based on entropy values, then change points in learning behavior log data are detected through a pruning precise timeline event algorithm, and meanwhile, Bayesian optimization is adopted to carry out iterative optimization on activity weight coefficients, transition weight coefficients and algorithm related parameters, so that the learning behavior log data are obtained. And finally, segmenting the continuous event log into sub-logs with significant behavior differences according to the detected change points, and importing the sub-logs into Cortado to realize log visualization. According to the method, comprehensive experiments are carried out on a real data set, so that the effectiveness of the method is proved, and the segmentation task of the learning behavior log can be accurately and efficiently completed.

Description

Technical Field

[0001] The present invention relates to the fields of learning behavior analysis and data processing, and in particular to a learning behavior log segmentation method based on behavior evolution. Background Art

[0002] With the growing development of the internet, online learning environments have become increasingly important in modern education. Online learning not only transcends the time and space constraints of traditional education but also provides learners with a wealth of learning resources and personalized learning recommendations. Learners' learning behaviors indirectly reflect their learning status (such as ability, interest, and motivation). This learning status evolves dynamically over time during the course of a course, influenced by multiple factors, including the learner's ability, the difficulty of the knowledge, and the learning environment. However, learners' learning behaviors during online learning are complex and dynamic, and the dynamic evolution of their behavior is a crucial indicator of learning changes and outcomes. Therefore, in-depth analysis of learners' behavioral processes in online learning environments has important theoretical and practical applications.

[0003] In this context, learning behavior logs, as a crucial data source for students' learning behavior, are a crucial basis for analyzing and understanding learners' learning processes. By analyzing learning behavior logs, educators can better understand learners' learning styles and patterns, thereby adjusting their teaching plans. However, most current learning behavior analyses treat the process represented by learning behavior logs as a simple continuous sequence, without further detailing the key turning points in the learning process—i.e., behavioral change points. Therefore, effective segmentation of learning behavior event logs is necessary to ensure consistency within log segments and variability between segments.

[0004] The core goal of log segmentation for learning behavior events is to transform unordered event streams into analyzable temporal or logical segments. These segmented segments can serve as the fundamental units for subsequent analysis, such as change comparison. While some existing log segmentation methods can initially segment behavioral data, these methods may overlook the evolutionary patterns inherent in learning behavior. Many previous log segmentation methods treat the entire learning process as a uniform flow of behavior, making it difficult to detect nonlinear behavioral changes and the locations of important turning points in the data. This results in the segmented log results failing to accurately reflect behavioral differences at different stages of a learner's life, thus compromising the accuracy of subsequent analysis. Furthermore, existing methods fail to fully account for the differences in information content and importance of different features, resulting in the failure of key features to fully play their due role in change point detection. Furthermore, most methods often overlook the correlations and interactions between features, failing to improve the overall log segmentation performance through systematic optimization. Therefore, developing a method that can automatically detect and accurately segment learning behavior based on its evolutionary characteristics is crucial for improving the accuracy and insight of learning behavior analysis. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a learning behavior log segmentation method based on behavior evolution with efficient algorithm and accurate segmentation.

[0006] The technical solution of the present invention to solve the above problem is: a learning behavior log segmentation method based on behavior evolution, comprising the following steps:

[0007] (1) Data preprocessing and feature extraction: convert the original learning behavior log obtained from the online teaching platform into an event log in XES format, define relevant features, and simultaneously extract activity type distribution, activity frequency temporal changes, transition type pattern, and event duration features; (2) Entropy weight allocation: calculate the information entropy values of activity frequency and transition frequency respectively, and dynamically allocate activity weight coefficients and transition weight coefficients based on the entropy values; (3) Change point detection: detect the mutation points in the behavior pattern based on the entropy weight allocation through the PELT (Pruned Exact Linear Time) algorithm, which is the pruning exact linear time algorithm; (4) Parameter joint optimization: use Bayesian optimization to iteratively optimize the activity weight coefficient, transition weight coefficient, and algorithm-related parameters until the relevant indicators reach the preset threshold or the maximum number of rounds; (5) Log segmentation output: segment the continuous event log into sub-logs with significant behavioral differences based on the detected change points, and import them into Cortado for log visualization.

[0008] Furthermore, in step (1), the raw log data extracted from the online learning environment is preprocessed and feature extracted, and the process includes the following steps:

[0009] (1) Clean and standardize the original log data of learning behavior, remove noise data and fill missing values;

[0010] (2) Convert the processed learning behavior data into event logs in XES format;

[0011] (3) Extract and define features from learning behavior log data;

[0012] (4) The learning behavior characteristics are normalized through statistical analysis and time series modeling to obtain a standardized feature matrix.

[0013] Furthermore, in step (2), weight distribution based on entropy value is performed, and the process includes the following steps:

[0014] (1) Using the activity frequency and transition frequency extracted from the original learning behavior log, the frequency of occurrence of each feature is calculated to obtain the feature probability distribution of the entire data;

[0015] (2) Calculate the entropy value that reflects the uncertainty and information content of each feature;

[0016] (3) In order to compare the entropy values of different features at the same scale, the entropy values are normalized;

[0017] (4) Map the entropy value to the weight coefficient of the feature, and obtain the weighted feature data for subsequent change point detection.

[0018] Furthermore, in step (3), based on the entropy weight distribution result obtained in step (2), the Bayesian optimization parameters of this round are input, and the change points in the learning behavior log are detected using the pruning exact linear time algorithm, thereby obtaining the basis for segmenting the log.

[0019] Furthermore, in step (4), Bayesian optimization is used to jointly optimize the activity weight coefficient, transition weight coefficient and algorithm-related parameters to minimize the objective function, obtain the best parameter setting, and obtain the optimal change point set, thereby further improving the accuracy and robustness of learning behavior log segmentation.

[0020] Furthermore, in step (5), after obtaining the final set of change points, ensure that each segment of data is processed correctly to ensure its integrity. At the same time, define the learning behavior event log after each segmentation to ensure its internal consistency and differences between segments. Finally, output the event log of each segment after processing and import it into Cortado to realize log visualization.

[0021] The beneficial effects of the present invention are:

[0022] This invention aims to effectively segment learning behavior event logs, improving the pruning exact linear time algorithm and using an entropy-weighted distribution method to weight features such as activity frequency and transition frequency. This ensures that features with greater information have a greater impact during change point detection in event log data. Furthermore, incorporating Bayesian optimization techniques, it iteratively adjusts activity weights, transition weights, and penalty parameters to maximize segmentation effectiveness, ensuring significant behavioral differences between different learning stages. After determining the final segmentation point, the learning behavior log is automatically divided into multiple sub-logs. This method improves the accuracy of learning behavior event log segmentation tasks, providing more refined and accurate data support for learning behavior analysis, pattern recognition, and personalized instruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is the overall framework diagram of the present invention;

[0024] Figure 2 Example of process tree for log visualization after segmentation;

[0025] Figure 3 Visualize BPMN model example for segmented logs;

[0026] Figure 4 This is an experimental comparison chart showing the impact of the number of Bayesian optimization iterations on the algorithm effect. DETAILED DESCRIPTION

[0027] Below in conjunction with accompanying drawing and example, the present invention is further described.Should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.In addition, should be understood that after reading the content of the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.

[0028] The overall framework of the learning behavior log segmentation method based on behavior evolution proposed in this invention is as follows: Figure 1As shown in the figure, the main purpose of this method is to segment continuous online learning behavior logs into multiple sub-logs by time period, facilitating subsequent analysis of learning behavior at different stages. First, data is preprocessed from the teaching system and converted into event logs in XES format. Activity type, number of activities, state change type, number of changes, and event intervals are then extracted from the logs. An entropy-based method is then used to calculate weights. Specifically, the entropy values of activities and state changes are calculated, and appropriate weights are assigned to each activity and state change. This ensures that informative features have greater influence when detecting change points. Next, Bayesian optimization techniques are used to iteratively optimize the activity weights, transition weights, and penalty parameters of the PELT algorithm to maximize the silhouette coefficient, resulting in more accurate and stable segmentation results. Finally, the optimized weights and parameters are used to accurately locate changes in learning behavior and segment the original log into multiple sub-logs with significant differences.

[0029] The following is a detailed description of a learning behavior log segmentation method based on behavior evolution proposed by the present invention, which includes the following steps:

[0030] (1) Data preprocessing and feature extraction: the original learning behavior logs obtained from the online teaching platform are converted into event logs in XES format, and relevant features are defined. Activity type distribution, activity frequency temporal changes, transition type patterns, and event duration features are simultaneously extracted.

[0031] Step (1) specifically includes the following steps:

[0032] Step (1.1) extracts the original learning behavior log data from the online teaching platform. The original data is a time series containing events, where each event represents a learning behavior, defined as E = {e1, e2, ..., e n}, where e i is the time of the i-th learning behavior, including attributes such as timestamp, activity type, and duration.

[0033] Step (1.2) cleans and standardizes the raw data, removes noise data and fills missing values, and obtains a standardized event data set D = {d1, d2, ..., d n}, where each data item d i A record of learning behavior that represents outcome processing.

[0034] Step (1.3) converts the processed data into an event log in XES format to obtain L XES ={l1,l2,...,l n}, where each l i An XES event log instance containing multi-dimensional information (activity type, timestamp, duration, etc.).

[0035] Step (1.4) extracts and defines the features in the log data, and defines the activity type distribution as A = {a1, a2, ..., a k}, where a i is the i-th activity type, and the activity frequency is The transition type is T={t1,t2,...,t l}, the transition frequency is Duration is

[0036] Step (1.5) normalizes the features by statistical analysis and time series modeling to obtain a standardized feature matrix X = {x1, x2, ..., x n}, where each x i It is the characteristic vector of the i-th event in the time series, which facilitates subsequent entropy calculation and change point detection.

[0037] (2) Entropy value weight allocation: calculate the information entropy values of activity frequency and change frequency respectively, and dynamically allocate activity weight coefficient and change weight coefficient based on the entropy values.

[0038] Step (2) specifically includes the following steps:

[0039] Step (2.1) First, perform statistical analysis and calculate the probability distribution using the activity frequencies extracted from the original learning behavior logs. and transition frequency Calculate the frequency of each feature. For each feature x i , the probability of its occurrence p(x i ) can be calculated by the following formula:

[0040]

[0041] Among them, count(x i ) represents the feature x i The number of occurrences in the learning log, where k is the total number of feature types. This calculation helps analyze the distribution of each activity and transition in the behavior log, providing a basis for subsequent entropy calculations. Probability distribution allows for quantitative analysis of the frequency of occurrence of each feature in the log, providing a basis for weighting.

[0042] After obtaining the probability distribution of the features in step (2.2), the entropy value H(X) of each feature is calculated. This value reflects the uncertainty and information content of the feature. The higher the entropy value, the more uncertain the distribution of the feature and the greater the information content. Conversely, the lower the entropy value, the smaller the information content of the feature. The formula for calculating the entropy value H(X) is as follows:

[0043]

[0044] Among them, p(x i ) is the feature x i The probability distribution of . Entropy calculation is used to measure the information content of each feature, thus providing a basis for subsequent weight allocation. By calculating entropy, important features in event log data can be effectively identified, allowing features with higher information content to have a greater impact on the log segmentation process.

[0045] In step (2.3), in order to compare the entropy values of different features at the same scale and avoid the influence of scale differences between features, the entropy value is normalized to fall within the interval [0,1]. The normalized entropy value H normalized (X) is calculated using the following formula:

[0046]

[0047] Among them, H max is the maximum entropy value, and the calculation formula is H max =log(k), where k is the total number of features. This represents the maximum entropy of the feature values, representing the situation when all features are evenly distributed. The purpose of entropy normalization is to unify the entropy values of different features into a standardized range to facilitate subsequent weighting and analysis.

[0048] Step (2.4) maps the entropy value to the weight coefficient w(x i ), this coefficient reflects the influence of the feature on the detection of change points in the learning behavior log. Each feature x i The weight of w(x i )=H normalized (x i ) assignment, which ensures that the features with higher information content have a greater influence in the subsequent change point detection, which helps to improve the accuracy and robustness of the log segmentation task. Finally, the weighted feature data X weigh ted Used for change point detection to identify important change points in learning behavior. By analyzing weighted features, we can effectively identify mutation points in learning behavior and then segment the learning behavior log into multiple sub-logs with significant differences. This process is crucial for subsequent behavioral difference analysis, helping to better understand the evolution of learners' behavioral patterns and provide more precise support for educational data analysis.

[0049] (3) Change point detection, based on entropy weight distribution, a pruning exact linear time algorithm is used to detect mutation points in behavior patterns.

[0050] Step (3) specifically includes the following steps:

[0051] Step (3.1) uses the weighted feature data X obtained in step (2.4) weighted , construct a time series dataset y of behavior logs 1:n ={y1,y2,...,y n},y i represents the i-th learning behavior event, such as "reading" and "practice", which has attributes such as timestamp and content. n represents the total number of events in the log. By detecting the change points in the log, the log is divided into m+1 sub-logs. Each log represents a stage of the learner's learning behavior. There is a high degree of similarity within the segment, but obvious differences between segments. τ 1:m is a set of change points, defined as τ 1:m ={τ1,τ2,...,τ m}.

[0052] In step (3.2) of the pruning exact linear time algorithm, in order to detect change points in the time series, it is necessary to calculate the cost of each possible change point. The cost function reflects the change in the behavioral characteristics of the time interval and consists of two main parts:

[0053] Cost function Measuring time interval [τ i-1 ,τ i The degree of data fit within the interval [ ], that is, the degree of aggregation of event features within this interval. If the behavioral pattern within this interval changes little, the cost of this part is low, and vice versa. The specific calculation method depends on how the time series is modeled and clustered.

[0054] Penalty term βm: This term is used to prevent over-segmentation of the time series. It encourages the model to reduce the number of change points. m is the number of change points, and β is the penalty coefficient related to the number of change points, which is used to balance the complexity and goodness of fit of the model.

[0055] Therefore, the cost function C(y 1:n ) is calculated as follows:

[0056]

[0057] in, Indicates the change point τ i-1 +1 to τ i The cost of this segment of data reflects the intensity of the behavioral change during that time period; m is the number of change points, and β is the penalty coefficient. The ultimate goal of the cost function is to find an appropriate number of change points so that the model can accurately identify significant changes in learning behavior while maintaining a reasonable level of complexity.

[0058] In step (3.3), to find the optimal change point, the pruning exact linear time algorithm combines the idea of dynamic programming to gradually determine the optimal split point by calculating the minimum cost at each time point. Assume that we have reached time point t, at which point the minimum cost F(t) of the previous t time points has been determined. For each potential change point τ, we calculate its cost and sum it up to obtain the minimum cost F(t) at the current time point. The formula is:

[0059]

[0060] Among them, F(t) represents the minimum cost from the beginning of the log to the time point t, that is, the cost of the optimal segmentation scheme when the log is divided into several stages from the beginning to t. To calculate the minimum cost to the time point t, we can select a previous moment τ as the segmentation point and add the optimal cost F(τ) of the previous time period to the log segmentation cost C(y from τ+1 to t) τ+1:t ) and the penalty term β are added together and minimized for all possible log split points to obtain the global optimal solution. At the same time, the pruning technique is used based on the "optimal substructure" property of the cost function: when a change point position τ is no longer optimal for the interval at the current time point t, it cannot become the optimal solution at any future time point. Certain candidate solutions that are unlikely to produce the optimal solution are eliminated, significantly optimizing computational efficiency. For example, if the behavior within a certain time period is highly consistent (e.g., all are "exercise"), then there is no need to insert additional split points in that segment.

[0061] (4) Joint parameter optimization: Bayesian optimization is used to iteratively optimize the activity weight coefficient, transition weight coefficient, and algorithm-related parameters until the relevant indicators reach a preset threshold or a maximum number of rounds.

[0062] Step (4) specifically includes the following steps:

[0063] Step (4.1) calculates the current silhouette coefficient (Silhouette Score), which combines the intra-cluster similarity (closeness) and inter-cluster difference (separation) to fully reflect the effect of log segmentation. A high silhouette coefficient indicates that the learning behavior within a log segment is highly similar, while the learning behavior between different log segments is significantly different. This is crucial for accurately capturing the evolution of learning behavior. The calculation process is as follows:

[0064]

[0065] Where a(i) is the average distance between sample i and the other samples in its cluster, indicating the compactness within the cluster; b(i) is the average distance between sample i and all samples in its nearest neighbor cluster, indicating the separation between clusters. The overall silhouette coefficient S is the average of the silhouette coefficients of all samples, N is the total number of samples, and s(i) is the silhouette coefficient of sample i.

[0066] Step (4.2) performs Bayesian optimization to maximize the silhouette coefficient of the segmented log segment. The goal of Bayesian optimization is to adjust the parameter vector θ = (α1, α2, ..., α n ,β1,β2,...,β m ,γ), to maximize the behavior log segmentation effect. Specifically, this paper uses the silhouette coefficient as the main evaluation indicator. Since Bayesian optimization is usually used for minimization problems, we transform the objective function into minimizing its negative value and define the objective function f(θ) as follows:

[0067] f(θ)=-Silhouette Score (8)

[0068] Among them, θ is the parameter set to be optimized, including activity weight coefficients α1, α2, ..., α n , transition weight coefficients β1,β2,...,β m and the penalty parameter γ in the exact linear-time algorithm for pruning.

[0069] In step (4.3), Bayesian optimization gradually updates the model through multiple iterations, selecting the most promising parameter combination for evaluation in each iteration. After each evaluation, the model is updated and the acquisition function selection strategy is adjusted. The optimization process continues until the objective function value reaches a preset threshold or the maximum number of iterations is reached. The optimal parameter combination obtained through multiple Bayesian optimizations is as follows:

[0070]

[0071] The optimized activity weight coefficient, transition weight coefficient and penalty parameter of the algorithm can effectively improve the accuracy of segmentation and ensure that the changes in learning behavior can be better captured. Based on this result, the change point detection in step (3) is performed again to obtain the final learning behavior change point, which is the optimal change point. In order to maximize the segmentation effect of the learning behavior log in the future.

[0072] (5) Log segmentation output: The continuous event log is segmented into sub-logs with significant behavioral differences based on the detected change points, and imported into Cortado for log visualization.

[0073] Step (5) specifically includes the following steps:

[0074] Step (5.1) converts the change points detected in step (4.3) As the basis for log segmentation, it is defined as each time point in the entire event log. In order to ensure that all data is processed, it is first necessary to ensure that the list of change time points includes the beginning and end of the log. If the last change point is not the end point of the log, end , then we need to add it to the list of change points. The final set of change points is:

[0075] T={t0,t1,t2,...,t n ,t end} (10)

[0076] Among them, t0=0 indicates the beginning of the log, t end Indicates the end time of the log. Then, based on these change points, the original log is divided into multiple sub-log segments.

[0077] Step (5.2) The original event log L = {l1,l2,...,l n} is divided into several sub-logs according to the change point L i , where each sub-log L i represents the time period [t i ,t i+1 ) in the learning behavior data. Each sub-log segment L i The corresponding time intervals are:

[0078] L i ={l i |t i ≤timestamp(l j ) <t i+1} (11)

[0079] Among them, l i It is an event in the original log, timestamp(l j ) is the timestamp of the event, indicating the specific time when the event occurred. Each sub-log L i It corresponds to a continuous time period and should have a consistent behavior pattern within that time period.

[0080] Step (5.3) Each sub-log will extract the corresponding event data from the original event log according to the time range defined in step (5.2). By extracting the data within each sub-log time period, a new sub-log segment is formed. Each sub-log segment L iEach sub-log represents a learning behavior pattern within a specific time interval. To facilitate subsequent analysis and processing, each sub-log will be saved as an independent XES file, which contains information about the event type, timestamp, and other related metadata. i We save it as a separate XES file and sequentially name it segment_{i+1}.xes to distinguish each sub-log. During the saving process, we ensure that each sub-log maintains the same structure as the original log, including event type, timestamp, and other relevant information, to ensure data consistency and integrity.

[0081] Step (5.4) imports the segmented log obtained in step (5.3) into the Cortado process mining tool, such as Figure 2 and 3 As shown, Cortado can represent each log segment using a process tree and BPMN model. The hierarchical structure of the process tree can decompose and simplify complex processes, making analysis more systematic and modular. This structured representation not only makes the model easier to understand but also facilitates automated analysis and optimization. The BPMN model graphically displays the control flow of each segment, allowing you to identify the sequence of activities, conditional branches, and simultaneous activities.

[0082] Example:

[0083] In this embodiment, an experimental verification is conducted on the learning behavior log segmentation method provided by the present invention. The data set and experimental settings, evaluation indicators, comparison methods and experimental results of this embodiment are described in detail below.

[0084] Dataset and experimental settings:

[0085] The present invention uses real-world datasets to conduct experiments to obtain solid experimental results. The dataset comes from the real learning behavior data recorded by students from the University of Genoa in an online learning environment. The specific content of the dataset includes course ID, learner ID, exercise content, activity name, start time, end time, etc. After preprocessing such as data cleaning, the final encoding of the relevant learning activities and activity names is shown in Table 1. After renaming, there are a total of 8 types of behavioral activities in the dataset. In the experiment, the learning behavior of the same learner in the course is analyzed each time, and the learning behavior log is segmented. The learning logs of different learners in the dataset are randomly selected. The effectiveness of the learning behavior log segmentation method proposed in the present invention is verified from three aspects: comparison with other methods, exploration of the impact of using different models on the method, and exploration of the impact of different optimization iteration numbers on the method.

[0086] Table 1 Learning activity coding table

[0087]

[0088] Evaluation indicators:

[0089] The purpose of the experiment is to prove that the algorithm has a good effect on the segmentation of learning behavior event logs, which requires good feature similarity within the segmentation results and certain differences between segments. In terms of local fitness, this experiment uses the silhouette coefficient (SS) as an evaluation method for the effect of different methods on the log segmentation task. The larger the value of the index, the better the log segmentation effect. Its calculation method is shown in formulas (6) and (7). In the evaluation of the global effect, the Davis-Bouldin Index (DBI) is used to evaluate the separation between all segments and the precision within each segment. The overall quality is obtained by combining the similarity of all clusters. The lower the DBI value, the better the segmentation effect. Its calculation method is shown in formula (12):

[0090]

[0091] Among them, S i is the degree of dispersion of samples within the i-th segment; d ij is the distance between the i-th segment and the j-th segment.

[0092] Comparison method:

[0093] In this section, we select two representative methods as baselines: Binary Segmentation, an excellent method for change point detection; Density-Based Spatial Clustering of Applications with Noise, a clustering algorithm widely used in various fields; and the traditional PELT algorithm. The methods are described as follows:

[0094] (1) BS (Binary Segmentation): A change point detection algorithm widely used in time series analysis, which recursively segments the time series into multiple parts with significant differences to identify change points in the data.

[0095] (2) DBSCAN (Density-Based Spatial Clustering of Applications with Noise): A density-based clustering algorithm that does not require a predefined number of clusters. It can be used to discover clusters of arbitrary shapes and effectively identify noise points. It can identify high-density areas (indicating stable stages of behavior) and low-density areas (possibly representing turning points or noise in behavior) in learning logs, thereby segmenting learning behavior logs.

[0096] Experimental results and analysis:

[0097] Table 2 shows the performance of different methods on the SS and DBI metrics. The improved PELT algorithm proposed in this paper outperforms the BS, DBSCAN, and original PELT algorithms in both the Silhouette Coefficient and DBI metrics, demonstrating that this algorithm can better maintain feature similarity in the learning behavior log segmentation task while making the distinction between different stages clearer. Taking Learner 1 as an example, the improved PELT algorithm improves the Silhouette Coefficient by 7.4%, 21.8%, and 11.5% compared to the BS, DBSCAN, and traditional PELT algorithms, respectively. This indicates that the improved algorithm achieves more consistent features across segmented paragraphs and greater differentiation between them. The DBI metric decreases by 51%, 74.5%, and 79.2%, respectively, demonstrating that the algorithm can better distinguish behaviors at different learning stages and reduce overlap and ambiguity between adjacent paragraphs. In terms of algorithmic principle, the traditional BS algorithm uses simple segmentation rules. When processing learning behavior logs, it may not accurately capture complex learning state changes, resulting in over-segmenting or missing important change points. The DBSCAN algorithm relies on density for clustering, and may fail to detect subtle changes in behavioral patterns when behavior changes slowly. The traditional PELT algorithm lacks sensitivity to gradually changing learning behaviors. The improved PELT algorithm in this paper incorporates entropy-weighted behavioral features and Bayesian optimization, taking into account the dynamic changes in learning behavior logs while maintaining global optimality. It fine-tunes changes in each data segment and uses adaptive data processing to more accurately locate behavior log segments.

[0098] Table 2 Performance comparison results with different baseline methods in log segmentation tasks

[0099]

[0100] Table 3 shows the impact of different cost functions used in the methods on segmentation performance. In traditional change point detection methods, the cost function is a measure of "homogeneity," and its choice determines the types of changes that can be detected. In the PELT algorithm, different cost functions are used to measure the "difference" between data segments and guide the algorithm in selecting appropriate change points during the segmentation process. The PELT algorithm aims to detect change points in the data by minimizing the overall error, so the choice of cost function directly affects the quality and performance of the segmentation results. The results show that the RBF cost function performs best and exhibits greater robustness for this task, followed by the Cosine cost function. Learner behavior data often exhibits highly nonlinear characteristics. For example, a learner may exhibit positive learning behavior at one stage and then suddenly shift to inefficient learning. Both RBF and Cosine are based on a positive semidefinite kernel function, which defines the similarity between data points. This kernel function maps the original data into a high-dimensional feature space, where nonlinear relationships in the data can be rendered linearly separable. Therefore, RBF and Cosine are suitable for processing nonlinear changes in learning behavior log segmentation tasks. They can measure the "similarity" between data points rather than direct differences. Compared with linear models or Euclidean distance-based measurement methods, they can more effectively capture the complex patterns and nonlinear changes in learner behavior, better adapt to the dynamics of learner behavior, and provide more accurate segmentation results.

[0101] The RBF cost function uses a Gaussian kernel function to effectively measure the similarity between data and capture nonlinear variations. The Cosine cost function, on the other hand, calculates the cosine similarity of two data points, evaluating the angular difference between them. This focuses on the direction of the data and may not fully capture the nonlinear characteristics of the data. When processing high-dimensional data, learning behavior logs may contain many dimensions such as student ID, activity type, change type, and time. The RBF cost function can handle high-dimensional data and capture complex relationships between different dimensions. However, cosine similarity may encounter the curse of dimensionality in high-dimensional spaces, resulting in less accurate similarity calculations.

[0102] Table 3 Effect of different cost functions used in the method on the segmentation effect

[0103]

[0104] Figure 4The impact of different optimization iteration counts on the method is shown in a line graph. The horizontal axis represents the number of optimization iterations, and the vertical axis represents the silhouette coefficient of the segmentation effect. By comparing the performance of the method at different iteration counts, we evaluated its impact on the stability and accuracy of the segmentation results. It is generally confirmed that, while taking computational cost into consideration, the method's performance stabilizes after around 100 iterations.

[0105] The present invention proposes an efficient and accurate learning behavior log segmentation method based on behavioral evolution. In the process of analyzing behavioral event logs, different features in complex learning behaviors may have different contributions and importances. This difference will have a great impact on the accuracy of the log segmentation task. To address this problem, the present invention calculates the entropy values of different features and converts them into corresponding feature weights, thereby improving the sensitivity of the method to changes in the learning stage. At the same time, Bayesian optimization is introduced to make up for the shortcomings of traditional methods in the parameter selection and optimization process, ensuring that the algorithm remains efficient and accurate when processing dynamic and complex data. The traditional pruning exact linear time algorithm is improved based on these two key points, improving the accuracy and robustness of the method in the entropy of learning behavior event log segmentation tasks.

Claims

1. A learning behavior log segmentation method based on behavior evolution, characterized in that: The following steps are involved: (1) Data preprocessing and feature extraction: convert the original learning behavior logs obtained from the online teaching platform into event logs in XES format, define relevant features, and simultaneously extract activity type distribution, activity frequency temporal changes, transition type patterns, and event duration features; (2) Entropy weight allocation: calculate the information entropy of activity frequency and change frequency respectively, and dynamically allocate activity weight coefficient and change weight coefficient based on the entropy value; (3) Change point detection: detect the mutation point in the behavior pattern through the PELT (Pruned Exact Linear Time) algorithm based on the entropy weight allocation; (4) Parameter joint optimization: use Bayesian optimization to iteratively optimize the activity weight coefficient, change weight coefficient and algorithm-related parameters until the relevant indicators reach the preset threshold or the maximum number of rounds; (5) Log segmentation output: segment the continuous event log into sub-logs with significant behavioral differences according to the detected change points, and import them into Cortado to realize log visualization.

2. The learning behavior log segmentation method based on behavior evolution according to claim 1 is characterized in that: In step (1), the raw log data extracted from the online learning environment is preprocessed and feature extracted, and the process includes the following steps: (1) Clean and standardize the original log data of learning behavior, remove noise data and fill missing values; (2) Convert the processed learning behavior data into event logs in XES format; (3) Extract and define features from learning behavior log data; (4) The learning behavior characteristics are normalized through statistical analysis and time series modeling to obtain a standardized feature matrix.

3. The learning behavior log segmentation method based on behavior evolution according to claim 1 is characterized in that: In step (2), weight distribution based on entropy value is performed, and the process includes the following steps: (1) Using the activity frequency and transition frequency extracted from the original learning behavior log, the frequency of occurrence of each feature is calculated to obtain the feature probability distribution of the entire data; (2) Calculate the entropy value that reflects the uncertainty and information content of each feature; (3) In order to compare the entropy values of different features at the same scale, the entropy values are normalized; (4) Map the entropy value to the weight coefficient of the feature, and obtain the weighted feature data for subsequent change point detection.

4. The learning behavior log segmentation method based on behavior evolution according to claim 1 is characterized in that: In step (3), based on the entropy weight distribution result obtained in step (2), the Bayesian optimization parameters of this round are input, and the change points in the learning behavior log are detected using the pruning exact linear time algorithm, thereby obtaining the basis for segmenting the log.

5. The learning behavior log segmentation method based on behavior evolution according to claim 1 is characterized in that: In step (4), Bayesian optimization is used to jointly optimize the activity weight coefficient, transition weight coefficient and algorithm-related parameters to minimize the objective function, obtain the best parameter setting, and obtain the optimal change point set, thereby further improving the accuracy and robustness of learning behavior log segmentation.

6. The learning behavior log segmentation method based on behavior evolution according to claim 1 is characterized in that: In step (5), after obtaining the final set of change points, ensure that each segment of data is processed correctly to ensure its integrity. At the same time, define the learning behavior event log after each segmentation to ensure its internal consistency and differences between segments. Finally, output the event log of each segment after processing and import it into Cortado to realize log visualization.

Images