A time-slot-based learner profiling and analysis method

Abstract

The application discloses a kind of learner portrait construction and analysis method based on time slot, first, based on course key event proposes adaptive time slot division method, the learning period between adjacent two key events is defined as a time slot.Second, construct individual time slot portrait LITP that can describe the dynamic change of learner interest and preference.Two methods for analyzing its sTP evolution are proposed, namely AsTP-DV and OsTP-EP.The former focuses on analyzing the degree of change between adjacent time slots sTP, and the latter focuses on analyzing the global evolution rule and trend of sTP.The relationship between learners is abstracted as a social network and combined with community detection algorithm to propose a time slot learning group division algorithm to divide learners into groups.The Top-N tags of representative nodes in the group are used as the basic tags of the group.Finally, the proposed method is verified using a real public dataset OULAD for multiple courses.The results show that the method is effective in constructing time slot dynamically changing learner portraits and time slot group portraits.

Description

Technical Field

[0001] This invention belongs to the field of educational data mining and learner modeling technology, and particularly relates to a method for constructing and analyzing learner profiles based on time slots. This method specifically encompasses adaptive time slot partitioning technology, dynamic modeling of individual time slot profiles, group segmentation based on complex network and social network analysis, and interpretability analysis that integrates temporal evolution patterns and group semantic labels. By introducing the concept of time slots to capture the dynamic evolution of learner preferences, and combining community detection algorithms and a multi-dimensional labeling system, it achieves multi-granular, temporal learner profile construction from individuals to groups, aiming to provide data-driven decision-making basis for precise teaching intervention and personalized learning support. Background Technology

[0002] The rapid development of big data technology and the widespread adoption of online learning have led to a continuous increase in the generation and storage of learner data, resulting in a series of problems such as "data explosion" and "information redundancy." How to extract valuable information from the vast amounts of log data on online learning platforms and construct learner profiles to help education administrators gain a more comprehensive understanding of learners and optimize the allocation of teaching resources has become a research hotspot in the field of educational data mining.

[0003] Traditional learner profiling research often focuses on static feature modeling, generating global labels through clustering or classification algorithms. Zhang Zhaowei et al., based on learning performance data from a data structure course, used hierarchical clustering to analyze students' ability distribution across different knowledge points, constructing a static student profile for exercise recommendation. Tian Yahui et al., based on data from multiple courses, designed a profile model with learners as the basic unit, summarizing five dimensions for learning analysis. Picado et al. proposed an online learning resource recommendation system, aiming to provide personalized learning resource recommendations for individual students through student profiling and keyword extraction techniques. However, such methods have three major limitations:

[0004] 1) Static profiles cannot capture the dynamic characteristics of learners' preferences as the course progresses, making it difficult to support real-time adjustments to teaching strategies.

[0005] 2) Most existing studies use individual learners as the smallest unit of analysis, resulting in coarse-grained profiles and a lack of fine-grained interpretability.

[0006] 3) The research perspective is too narrow, neglecting the interaction between individuals and groups.

[0007] To address the aforementioned problems, this invention proposes a time-slot-based learner profiling construction and analysis method, which is validated on multiple courses using real publicly available datasets. Experimental results demonstrate the effectiveness of the method. The main contributions of this invention are as follows:

[0008] 1) A new perspective for studying learner profiles is proposed, a time-slot-based learner profile is constructed, and an adaptive time-slot division method is designed to describe the dynamic changes of learner preferences and interests with multiple continuous sub-time-slot profiles, thereby finely depicting the evolution process of learner profiles.

[0009] 2) Using two fine-grained analysis methods, AsTP-DV and OsTP-EP, we conducted an in-depth study on the evolution characteristics of sub-slot profiles and their components.

[0010] 3) An Intra-Slot Learning Group Partitioning (ITLGD) algorithm is proposed, which utilizes similarity metrics to construct a learner social network and combines it with the Louvain community detection algorithm to achieve high-cohesion group partitioning. Furthermore, a Simple Multidimensional Labeling System (SMLS) is designed to generate semantic group labels based on the Top-N features of representative nodes, thereby improving the interpretability of the groups. Summary of the Invention

[0011] This invention provides a time-slot-based dynamic construction and analysis method for learner profiles. It introduces an adaptive time-slot division mechanism to capture the evolutionary patterns of learner behavior as the course progresses, and proposes an individual time-slot profile modeling method that integrates learning resources, activity types, and learning behavior preferences to achieve fine-grained characterization of learner states. Based on this, the invention proposes two time-series analysis methods, AsTP-DV and OsTP-EP, to quantify the dynamic evolutionary characteristics of learner profiles from two dimensions: local fluctuations and global patterns, respectively. Furthermore, by constructing a learner similarity network and combining it with the Louvain community detection algorithm, high cohesion of learning groups within time slots is achieved, and a simple multi-dimensional labeling system is designed to generate semantic and interpretable comprehensive labels for groups, effectively enhancing the pedagogical operability of the profile results. Experiments on the public educational dataset OULAD verify the effectiveness of the proposed method in characterizing learner dynamic behavior, identifying group structure, and analyzing evolutionary trends. Compared with existing static profile methods, it shows significant improvements in dynamism, interpretability, and group correlation.

[0012] To achieve the above objectives, the present invention adopts the following technical solution:

[0013] A method for constructing and analyzing learner profiles based on time slots includes the following steps:

[0014] Step 1: Data Preprocessing and Adaptive Time Slot Division. Raw log data of learners is obtained from online learning platforms (such as the OULAD dataset), including records of resource clicks, activity participation, quiz submissions, and demographic information. After data cleaning and feature extraction, an adaptive time slot division method is proposed: using the deadline of a "weighted quiz"—a time-unique quiz strictly synchronized with the teaching progress—as the key node, the entire learning cycle is divided into N consecutive time slots. Each time slot covers the complete learning cycle between two adjacent weighted quizzes, thus providing an ideal time scale for subsequent dynamic analysis.

[0015] Step 2: Construct individual learner time-slot profiles. For each time slot defined in Step 1, construct a sub-time-slot profile for each learner. This profile is organized in a tree structure and includes three hierarchical components: learning resource preference: calculating activity level based on indicators such as the relative total number of clicks and interaction days for resources; activity type preference: calculating the preference ratio of different activity categories (such as forums and quizzes) based on the interaction with the associated resources; and learning behavior preference: aggregating and calculating the distribution of different learning behaviors (such as previewing, knowledge acquisition, interaction, and consolidation) based on activity type preferences. In addition, the profile of the first time slot also includes the learner's static demographic information. The complete individual time-slot profile of a learner throughout the entire course is composed of the sub-time-slot profiles of all their time slots in sequence.

[0016] Step 3: Analyze the dynamic evolution of individual time-slot profiles. To deeply analyze the dynamic changes in learners' preferences, two complementary analytical methods are proposed: AsTP-DV (Analysis of the degree of change between adjacent time slots) and OsTP-EP (Analysis of the overall evolutionary pattern). The former calculates the difference in preference values ​​of each component (Ma, Ca, Be) of the sTP between adjacent time slots, and quantifies the local volatility and stability of learners' preferences by calculating the entropy value of the probability distribution of these differences; the latter treats the activity sequence of each component in N time slots as a whole, and identifies three evolutionary sub-patterns: "stable," "rising," and "falling" by defining "evolutionary segments" and setting thresholds. Furthermore, combining high / low activity states, six typical global evolutionary patterns are summarized, such as "high-activity stable type," "high-activity transitional type," and "occasional participation type."

[0017] Step 4: Construct a learner social network and calculate similarity within a time slot. For a single time slot, learners are abstracted as network nodes. By comprehensively calculating the similarity of interaction behaviors, performance, and preferences among learners, and obtaining the total similarity through weighted summation, a weighted learner social network is constructed.

[0018] Step 5: Implement intra-slot learning group segmentation. Based on the social network constructed in Step 4, an intra-slot learning group segmentation algorithm is proposed. This algorithm first divides learners into three similarity groups (high, medium, and low) according to a similarity threshold. For the high and medium similarity groups, the graph structure is initialized, and the Louvain community detection algorithm is applied to optimize modularity, performing multi-level community segmentation to identify highly cohesive secondary learning groups. Learners in the low similarity group are considered as independent individual groups.

[0019] Step 6: Generate Group Profiles and Interpretable Tags. This step constructs a time-slot-specific group profile for each learning group identified in Step 5. First, the most representative core learners are identified by calculating the eigenvector centrality of nodes within the group. Then, the Top-N learning resources, activity types, learning behaviors, and key achievements of these representative nodes are extracted as basic group tags. Finally, a simplified multi-dimensional tagging system is designed to integrate semantic tags from three dimensions: behavioral patterns, academic performance, and homework habits, generating comprehensive group tags such as "Collaborative Preparer - Excellent - Early Submission," greatly enhancing the interpretability of the group profile.

[0020] Step 7: Method Validation and Case Analysis. The method is systematically validated on real-world public datasets (such as multiple courses in OULAD). By showcasing the dynamic LITP of specific learners, the evolutionary entropy plot of their sTP, the overall pattern diagram, and the group segmentation results and comprehensive labels for different time slots, the effectiveness and superiority of this method in capturing learner dynamic behavior, segmenting meaningful learning groups, and providing interpretable insights are qualitatively and quantitatively demonstrated. Attached Figure Description

[0021] Table 1 lists the abbreviations of the terms used in this invention;

[0022] Figure 1 This is a general framework diagram of the technical solution of the present invention;

[0023] Figure 2 A schematic diagram of a sub-slot image for a single time slot;

[0024] Figure 3 A schematic diagram of individual learner time slot profiles;

[0025] Figure 4 A schematic diagram of AsTP-DV;

[0026] Figure 5 A schematic diagram of OsTP-EP;

[0027] Figure 6 There are three evolutionary sub-patterns;

[0028] Figure 7Individual time-slot profiles for different learners in different courses; (a) Individual time-slot profile of learner A in course 1; (b) Individual time-slot profile of learner B in course 2;

[0029] Figure 8 Entropy graphs for learners A and B corresponding to the components of sTP; (a) Entropy graph of the difference in Ma of A; (b) Entropy graph of the difference in Ca of A; (c) Entropy graph of the difference in Be of A; (d) Entropy graph of the difference in Ma of B; (e) Entropy graph of the difference in Ca of B; (f) Entropy graph of the difference in Be of B.

[0030] Figure 9 The overall evolution pattern of activity type preferences in the sTP components of learner A: (a) High-activity stable type H, SP; (b) High-activity transient type CT; (c) High-activity fluctuating type F, Q; (d) Low-activity stable type RS; (e) Low-activity fluctuating type U; (f) Occasional participation type DPS, QN;

[0031] Figure 10 The overall evolution pattern of activity type preferences in the sTP components of learner B: (a) high-activity volatile type CT, SP; (b) high-activity stable type H; (c) high-activity transient type F; (d) low-activity volatile type WK, U; (e) low-activity stable type RS; (f) occasional participation type G, CB;

[0032] Figure 11 The similarity matrix diagrams for time slot 1 of Course 1 and time slot 7 of Course 2 are as follows: (a) High similarity matrix for time slot 1 of Course 1; (b) Medium similarity matrix for time slot 1 of Course 1; (c) Low similarity matrix for time slot 1 of Course 1; (d) High similarity matrix for time slot 7 of Course 2; (e) Medium similarity matrix for time slot 7 of Course 2; (f) Low similarity matrix for time slot 7 of Course 2.

[0033] Figure 12 The learning groups are designated as 1-0-0 for time slot 1 of course 1 and 7-0-4 for time slot 7 of course 2. Detailed Implementation

[0034] like Figure 1 As shown, this invention provides a method for constructing and analyzing learner profiles based on time slots, including the following steps:

[0035] Step 1: Data Preprocessing and Adaptive Time Slot Division. Raw log data of learners is obtained from online learning platforms (such as the OULAD dataset), including records of resource clicks, activity participation, quiz submissions, and demographic information. After data cleaning and feature extraction, an adaptive time slot division method is proposed: using the deadline of a "weighted quiz"—a time-unique quiz strictly synchronized with the teaching progress—as the key node, the entire learning cycle is divided into N consecutive time slots. Each time slot covers the complete learning cycle between two adjacent weighted quizzes, thus providing an ideal time scale for subsequent dynamic analysis.

[0036] Step 2: Construct individual learner time-slot profiles. For each time slot defined in Step 1, construct a sub-time-slot profile for each learner. This profile is organized in a tree structure and includes three hierarchical components: learning resource preference: calculating activity level based on indicators such as the relative total number of clicks and interaction days for resources; activity type preference: calculating the preference ratio of different activity categories (such as forums and quizzes) based on the interaction with the associated resources; and learning behavior preference: aggregating and calculating the distribution of different learning behaviors (such as previewing, knowledge acquisition, interaction, and consolidation) based on activity type preferences. In addition, the profile of the first time slot also includes the learner's static demographic information. The complete individual time-slot profile of a learner throughout the entire course is composed of the sub-time-slot profiles of all their time slots in sequence.

[0037] For N time slots A certain time slot For example, the sub-slot profile of learner S is denoted as in, They represent time slots respectively. The start and end times; This indicates that the student's basic information consists of m features; Ma, Ca, and Be are denoted as: , indicating that there are n learning resources in this time slot; , indicating that there are a total of p types of activities; Let q represent the number of learning behaviors. Ultimately, the individual time-slot profile of learner S is denoted as . .

[0038] To quantitatively describe learners' preferences for learning resources, activity types, and learning behaviors across different time slots, this invention quantifies learners' behavioral characteristics and learning habits by calculating the activity level of each item. For a learning resource... Its activity level is calculated from four perspectives, using the formula. The calculation shows that, The percentage of the total number of clicks for the resource. The percentage of relative interaction days for resources The relative highest number of clicks for the resource and The longest consecutive number of days the resource has been used. The first three items are respectively derived from the formulas... Calculations show that This was derived from statistics. These are the corresponding weights, and . and They will all be normalized.

[0039]

[0040]

[0041]

[0042]

[0043] in, Representing resources The number of days that occur within a time slot; express Total number of clicks on day j; Indicates time slot The number of days.

[0044] Each activity type has a percentage relative to other activity types in the same time slot. The percentage is determined by the activity category. The relative number of days of resource usage and the relative total number of clicks are determined as shown in the formula. As shown.

[0045]

[0046] in, Indicates that it belongs to the activity type A collection of learning resources Indicates learning resources In the time slot The number of days of interaction between China and the United States Indicates students' interest in learning resources The total number of clicks on day j of its interaction period. These are the corresponding weights, and their sum is 1.

[0047] For learning behavior, the sum of behavioral preferences across time slots is 1. (Learning behavior) From the formula The calculation yielded the result.

[0048]

[0049] in, This indicates that it belongs to learning behavior. A set of activity types.

[0050] After the above steps, a sub-slot profile of the learner can be constructed. By repeating this process N times, a complete individual slot profile of the learner can be obtained.

[0051] Step 3: Analyze the dynamic evolution of individual time-slot profiles. AsTP-DV focuses on the differences in sub-time-slot profiles between adjacent time slots, aiming to assess the stability of learners' preferences across different time slots by quantifying the probability distribution of these differences and then calculating entropy values. Specifically, it first extracts the learner's sub-time-slot profile member preferences from time slot 0 to time slot N from the individual time-slot profile. In particular, time slot 0 is defined as the time period from course release to the start of the course. During this stage, the platform still records learner interaction data with the course, so the data in time slot 0 can be regarded as learners' pre-study behavior. The learner's sub-time-slot profile consists of three parts, with adjacent time slots... The difference in sub-slot profiles is defined as follows: ,in These represent the differences in preferences for learning resources, activity types, and learning behaviors across adjacent time slots. The following explanation uses activity type preference as an example. For N+1 time slots... The activity type set is In time slots In the context, the set of activity types is ,Then, ,in express Time slot Activity level minus Time slot The activity level. Thus, the N+1 time slots have a total of indivual Then, for each pair of adjacent time slots in the N+1 time slots, the probability distribution of the difference between each pair of time slots is first calculated, and then the stability of these distributions is described by the entropy value. This entropy value can reflect the degree of fluctuation in the changes between time slots. Specifically, the larger the entropy value, the more significant the change in the learner's preference between time slots; the smaller the entropy value, the more stable the change in the learner's preference between time slots.

[0052] OsTP-EP quantifies the analysis of learner inter-slot evolution patterns by analyzing the evolution of a series of numerical sequences. These numerical sequences refer to the activity sequences composed of various preferences across N time slots. Through a comprehensive examination of different sequences, this invention finds that every two nodes can effectively delineate the smallest evolutionary segment, which is also the most concise representation. This method is illustrated using learner activity type preferences as an example. This invention defines different evolutionary segments based on three evolutionary sub-patterns. ,in IV and h represent the initial value of the evolutionary segment and the difference between the two nodes, respectively.

[0053] However, the activity level changes very little, such as Since the sub-pattern remains unchanged before and after, this project defines a threshold after sufficient experimentation. =0.01 to distinguish evolutionary subpatterns: If a certain EF in Then the evolutionary segment ;like ,but ;like ,but .

[0054] For a specific type of activity ,in Indicates a specific activity type In the time slot The level of activity. Now, It can be represented by a sequence of N-1 evolutionary segments. The proportion can be expressed by the formula The calculation yielded the result.

[0055]

[0056] In order to understand In After extensive experimentation, dynamic thresholds were set for different activity types to categorize activity levels into high and low activity levels. ,but This time slot represents a high-activity type of activity. Conversely, activities with low activity levels are considered low-activity activities. , The proportion can be expressed by the formula The calculation yielded the result.

[0057]

[0058] After conducting experiments on the overall evolution patterns of time slot preferences among a large number of different learners, this invention summarizes six overall evolution patterns: high-activity stable, high-activity fluctuating, high-activity migration, low-activity stable, low-activity fluctuating, and occasional participation.

[0059] Highly active and stable: If satisfy Conditions, then The overall evolutionary pattern is classified as high-activity stable, indicating that learners have a high level of interest in this specific type of activity and their activity level remains relatively stable.

[0060] High-activity fluctuation type: if satisfy Conditions, then The overall evolutionary pattern is classified as high-activity fluctuation type, indicating that learners have a high level of interest in this specific type of activity, but the range of change fluctuates greatly.

[0061] Highly active migration: If satisfy Conditions, then The overall evolutionary pattern is classified as a high-activity transition, indicating that learners have a high interest in this specific type of activity, but there is a significant change in activity level before and after a certain time slot.

[0062] Low-activity stable type: if satisfy Conditions, then The overall evolutionary pattern is classified as low-activity stable, indicating that learners have low interest in this particular type of activity but maintain a relatively stable level of activity.

[0063] Low-activity fluctuation type: if satisfy Conditions, then The overall evolutionary pattern is classified as low-activity fluctuation type, indicating that learners have low interest in this specific type of activity and the fluctuation range is large.

[0064] Occasional participation type: if satisfy Conditions, then The overall evolutionary pattern is classified as occasional participation, indicating that learners have little interest in this particular type of activity and only interact occasionally due to task requirements.

[0065] Step 4: Constructing a learner social network and calculating similarity within a time slot. Within a given time slot, not only are sub-time slot profiles of all learners collected, but also information on learners' interactive behaviors and related test scores within that time slot. Based on this, the present invention comprehensively considers the similarity of learners' preferences, test scores, and interactive behaviors. Specifically, the total similarity between learners x and y in time slot i is defined as the weighted sum of the three similarities: interactive behavior similarity. Similarity of scores Similarity to preferences , as in the formula As shown, where .

[0066]

[0067] Similarity of interactive behaviors This is calculated by analyzing learners' interaction behavior with learning resources within time slots. First, a... The matrix records the actions of all learners Ls in time slot i within that time slot. The number of clicks on the learning resource Let represent the set of learning resources that learner j has interacted with in time slot i, and let the values ​​of the matrix represent the number of times each learner accesses a learning resource. Then, cosine similarity is used to calculate the similarity of the interaction behaviors between any two learners.

[0068] similarity of scores This is calculated by analyzing the learners' test score vectors. Each learner maintains a score vector of length 5.

[0069]

[0070] Where numAsmts represents the number of tests; TMAAsmtScore represents the score of the weighted test; TMAAsmtSubDate represents the submission date of the weighted test (expressed as the deadline minus the submission date); avgCMAAsmtScore represents the average score of the unweighted test; and avgCMASubDate represents the average submission date of the unweighted test. The similarity of learners' scores is also calculated using cosine similarity.

[0071] Preference similarity This reflects learners' interests and preferences regarding learning resources and activity types. Each learner maintains a database of length [length missing]. The preference vector, where This represents the number of learning resources that learner j interacted with within time slot i. This represents the number of activity types that learner j participated in within time slot i. The values ​​of the vector represent the learner's activity level towards learning resources and activity types. The preference similarity between learners is calculated using cosine similarity.

[0072] Extensive experiments have demonstrated that the distribution of learner similarity scores largely conforms to a normal distribution. To differentiate the degree of closeness between learners, this invention... Using a threshold, learner similarity is divided into three groups: high, medium, and low, thus forming three different similarity matrices.

[0073] High similarity groups are those with higher similarity within time slots. Learners in a high-similarity group are considered as a graph, where nodes represent learners and edges represent the similarity between learners. This graph may contain one or more connected components, each potentially containing more closely connected subgroups. Learners in high-similarity groups have distinct characteristics, strong connections within each group, and significant differences between groups.

[0074] Within the medium similarity group, the similarity within the time slot is between and Learners in the medium similarity group. These learners have some connection, but the degree of connection is lower than that in the high similarity group. Learners in the medium similarity group have relatively weak characteristics, and the differences between groups are small. Typically, these learners complete learning tasks methodically, lack obvious individual characteristics, and represent the average learner in the course.

[0075] Low similarity groups are those with similarity levels below a certain threshold within a time slot. These learners have almost no contact with other learners and can be considered an isolated group of individuals. However, it is clear that the probability of such learners appearing in pre-processed online courses is extremely low.

[0076] Step 5: Implement Intra-Slot Learning Group Partitioning. To further explore potential secondary groups within similarity groups, this invention proposes an Intra-Slot Learning Group Partitioning (ITLGD) algorithm, combining the widely used Louvain community detection algorithm, as shown in Algorithm 1. First, an empty graph is initialized for each similarity group, and then filled in sequentially from high to low similarity between learners. For high similarity groups, learners are added to the graph as nodes, and similarity values ​​are added as edges. Since the graph may contain multiple connected components, a modularity threshold is set. For connected components with a modularity less than this threshold, the partitioning is canceled, and the entire connected component is treated as a single group and added to the learning group; for connected components with a modularity greater than or equal to the threshold, each community within it is partitioned into an independent group. Second, when filling in the intermediate similarity groups, it is ensured that learners already belonging to the high similarity group are not added repeatedly, and the same partitioning method as for the high similarity group is used for group construction. Finally, for low similarity groups, ensuring that learners from the first two groups are not included, each learner in the low similarity group is treated as an independent group.

[0077] Regarding the algorithm output In order to gain a deeper understanding of the common and prominent characteristics of learners within a learning group, this invention identifies the most representative learners in the group by calculating the eigenvector centrality, and extracts the top-5 learning resources with the most interactions, the top-3 activity types with the highest activity levels, the top-2 learning behaviors, and weighted test scores of each learner as basic group label information.

[0078] At once In this context, the learning group includes learners with moderate similarity within a time slot, indicating that learners have certain similarities in terms of behavioral preferences and performance, but these similarities are not prominent. Therefore, this group can be considered as ordinary learners in the classroom.

[0079] At once In this regard, the learning group, namely learners who have deviated from the normal progress of the course, needs separate guidance from education administrators.

[0080]

[0081] Step 6: Generate group profiles and interpretable tags. To further enhance the interpretability of the learning group and comprehensively depict its behavioral patterns and learning outcomes, this invention proposes a multi-dimensional tagging system based on basic group tagging information, combined with behavioral characteristics, academic performance, and homework habits. This system is built upon three core dimensions: behavioral patterns, academic performance, and homework submission habits, ultimately forming interpretable semantic tags to support instructional strategy design and personalized intervention.

[0082] As shown in step 2, there is a hierarchical relationship between learning behaviors, activity types, and learning resources, with granularity ranging from coarse to fine. Considering the generalizability of this tagging system across different courses, this invention takes the coarsest-grained learning behavior as the behavioral feature.

[0083] The online learning process mainly includes the preparation stage, the acquisition stage, the interactive reflection stage, and the consolidation stage. The preparation stage is the learner's preparatory process before formally beginning online learning; the acquisition stage is the most important online learning process and the initial stage where learners acquire knowledge; the interactive reflection stage involves learners interacting with teachers and peers and reflecting on their own interactions; and the consolidation stage is the process where learners consolidate and internalize knowledge. After numerous experiments, six labels for behavioral characteristics were summarized, as shown in Table 2.

[0084] Based on the course grading criteria and the relative relationship between the submission time and deadline for key quizzes, the performance dimension and the assignment submission habit dimension are each divided into three levels:

[0085]

[0086] Based on the above three dimensions, the final learning group tags are generated, in the following format:

[0087]

[0088] Step 7: Method Validation and Case Analysis.

[0089] 1) To verify the feasibility and generalizability of the proposed method, this invention uses the publicly available real-world dataset OULAD. This dataset contains information on 22 courses and 32,593 learners, their in-class quiz results, and their logs of interaction with the Virtual Learning Environment (VLE). This invention selects two courses from the 22 courses as research cases, and after data preprocessing, the following data is obtained:

[0090] a) Course FFF2013J, hereinafter referred to as Course 1, included 991 learners. Each learner completed all 12 in-class quizzes, including 6 authorized and 6 unauthorized quizzes. The course recorded 860,132 interaction log entries, including 4 learning behaviors, 16 activity types, and 526 learning resources.

[0091] b) Course DDD2013J, hereinafter referred to as Course 2, included 684 learners, each of whom completed all 7 weighted in-class quizzes. The course recorded 423,174 interaction logs, including 4 learning behaviors, 10 activity types, and 456 learning resources.

[0092] The correspondence between learning behaviors and activity types is shown in Table 3, and the meaning of activity types is shown in Table 4. Each activity type corresponds to multiple learning resources.

[0093] 2) To construct individual learner time-slot profiles, this invention proposes a method for constructing individual time-slot profiles that can describe in detail the learner's preferences and interests in specific time slots and demonstrate the dynamic changes between time slots. The individual learner time-slot profile includes a corresponding number of sub-time-slot profiles, each represented in a tree structure, consisting of three members: the learner's learning resource preferences, activity type preferences, and learning behavior preferences. Furthermore, each learner's first time slot also includes their basic static information.

[0094] a) Time Slot Allocation: To determine a reasonable number of time slots, this invention uses the deadline for authorized tests as the basis for allocation because these slots have the characteristics of gradually opening as the course progresses and being unique within each slot. Therefore, Course 1 ultimately has 6 time slots, and Course 2 has 7 time slots.

[0095] b) Other parameters: After extensive experimental testing, this invention has determined...

[0096] Individual time-slot profiles, such as Figure 7 As shown. The individual time-slot profile of learner A from Course 1 is as follows. Figure 7 As shown in (a), the individual time-slot profile of learner B from Course 2 is as follows: Figure 7 As shown in (b).

[0097] In addition, the following provisions apply:

[0098] a) A black mark indicates that the content is appearing for the first time; a red mark indicates that the content's activity level has increased compared to the previous time slot; a blue mark indicates that the content's activity level has decreased compared to the previous time slot.

[0099] b) Learning resources are represented by a lowercase activity type code plus a serial number. For example, the 10th resource on the homepage is represented as h_10.

[0100] from Figure 7 From .a, we know that learner A is between 0 and 35 years old, male, a junior high school graduate, from the West Midlands, in an area with an economic level of 40-50%, and is not disabled. It is evident that this learner's learning resources, activity types, and learning activity levels varied across all six time slots. Compared to... and Regarding the types of learning resources and activities, learners may no longer access some resources / activities, while they may access some new ones. As for learning behavior, activity levels have changed, but remain in a relatively balanced state.

[0101] exist Figure 7 In .b, learner B's basic information is clearly available. This learner's learning resources, activity types, and learning activity levels all change across the seven time slots. Compared to... Learner B in The activity level of the learning behavior Preparation increased significantly, while the activity levels of other learning behaviors decreased significantly. It can be inferred that the learner's focus during different learning time slots also changed.

[0102] 3) Based on the constructed individual time-slot profile, this section focuses on the evolution of sub-time-slot profiles (sTPs) within the learner's individual time-slot profile (LITP). For a specific learner, their sTP comprises numerous components, such as preferences for certain learning resources, activity types, or learning behaviors. Since these components can all be represented numerically, this invention employs a unified method for their analysis. Through the evolutionary analysis of each sTP component, this invention ultimately achieves a comprehensive analysis of the overall evolution of the learner's individual time-slot profile (LITP). The following will use learners A and B as examples to specifically demonstrate their evolutionary analysis process.

[0103] For the AsTP-DV method, this invention first calculates the differences between the sTP components of adjacent time slots, then calculates the probability distribution of these differences in the corresponding time slots, and finally uses the entropy formula to calculate the entropy of the differences. It is worth mentioning that this method introduces the concept of time slot 0, i.e., the time slot before the course starts. The dataset centrally records log information from the 18 days before the course starts, which can effectively reflect learners' pre-course preparation. A higher entropy value indicates more drastic changes in the sTP components corresponding to a learner before and after the time slot, and vice versa. The entropy value graphs of the sTP components corresponding to learners A and B are shown below. Figure 8 As shown.

[0104] exist Figure 8In the diagram, 8.a-8.c are the entropy maps of the sTP components for learner A, and 8.d-8.f are the entropy maps of the sTP components for learner B. It is evident that the three entropy maps for learners A and B all show a progression from relatively high entropy values ​​for differences in learning resources to relatively low entropy values ​​for differences in learning behaviors. This is because the number of sTP components Ma, Ca, and Be decreases, and their granularity decreases, resulting in a decrease in the total entropy value of the three components. Therefore, it is reasonable for the relative entropy values ​​to decrease from large to small. From... Figure 8 From .b, we know that learner A's Ca difference entropy is in arrive The changes during this period were minimal compared to other time slots, indicating that learner A's activity type preferences remained relatively stable with minimal variation. Through AsTP-DV, education administrators can clearly understand the extent of learner changes as the course progresses and take timely intervention measures for any anomalies.

[0105] The OsTP-EP method primarily focuses on the overall evolution of various sTPs within the LITP. For each component, a numerical sequence consisting of three seed evolution patterns with different parameter values ​​can represent the corresponding component. Next, this invention will elaborate on this method using the sTP component activity type preferences of learners A and B as examples.

[0106] Based on the above experiments, this invention has obtained six overall evolutionary patterns of activity types. The representative overall evolutionary patterns of activity type preferences for learners A and B are as follows: Figure 9 , Figure 10 As shown in the image. The vertical axis represents the activity of Ca, and the horizontal axis represents the time slot. For high-activity stable types, see... Figure 9 a、 Figure 10 .b, High-activity fluctuation type Figure 9 .c、 Figure 10 .a, Highly active migration patterns are seen Figure 9 .b、 Figure 10 .c, Low-activity stable type see Figure 9 .d、 Figure 10 .e, Low-activity fluctuation type Figure 9 .e、 Figure 10 .d, occasional participation type Figure 9 .f、 Figure 10 .f.

[0107] from Figure 9 a、 Figure 10 As shown in .b, high activity in time slots accounts for a significant proportion of these components, and the sub-evolutionary pattern... The proportion of activity is dominant, and the activity difference h between adjacent time slots is within a very small range.

[0108] exist Figure 9 .c、 Figure 10 In .a, similarly, high activity in time slots dominates, but The proportion is very small and the range of h is very large.

[0109] exist Figure 9 .b、 Figure 10 In .c, high activity in time slots also accounts for a relatively high proportion, and the various sub-evolutionary modes are relatively balanced, but there are cases where h is particularly large before and after the time slot.

[0110] exist Figure 9 .d、 Figure 10 In .e, low activity in time slots accounts for a large proportion. It is dominant and there exists a value of h greater than a specific value.

[0111] exist Figure 9 .e、 Figure 10 In .d, similarly, low activity in time slots dominates, but The proportion is very small, and h also falls into a small range and is not all zero.

[0112] exist Figure 9 .f、 Figure 10 In .f, it is obvious that The ratio is 1, and the range of h is very small, with many cases where it is 0.

[0113] In summary, the six holistic evolution scenarios proposed in this invention can cover all possible holistic evolution scenarios except for chaotic evolutionary data. Furthermore, these experiments demonstrate the effectiveness of the agreed-upon rules in classifying a holistic evolution. In the experiments of this invention, The proportional threshold and the range of h can be summarized, but they may not work for other scenarios and require further research.

[0114] 4) Based on the time slots divided by the data preprocessing module, this section focuses on constructing and analyzing learner group profiles within a single time slot using real-world examples. For a single time slot, not only are all learners' STPs included, but also their performance information. This invention calculates learner similarity within a time slot, abstracts a learner social network, transforms the learner classification problem into a community detection problem within the social network, and then assigns group labels based on prominent nodes in the divided group to complete the profile interpretation. The following will use time slot 1 of Course 1 and time slot 7 of Course 2 as examples to specifically demonstrate the construction and analysis process.

[0115] a) Learner similarity: After extensive experimental verification, the formula... middle Course 1, Slot 1 Time slot 7 of course 2 Time slot 1 of course 1 and time slot 7 of course 2 pass the threshold. A schematic diagram of the similarity matrix after partitioning is shown below. Figure 11 As shown, nodes represent learners, and edges represent learner similarity.

[0116] exist Figure 11 In the diagram, ac represents the similarity matrix of time slot 1 in course 1; df represents the similarity matrix of time slot 7 in course 2. Figure 11 In .a and 11.d, the highly similar matrix graphs all have a large number of connected components, and most of these connected components have few nodes, sometimes only two. This indicates that learners with high similarity are relatively dispersed within time slots, and that learners encounter relatively few people with high similarity. Figure 11 In .b and 11.e, the medium similarity matrix graphs have only one connected component, and the density of this connected component is greater than 96%. This indicates that each learner in the medium similarity matrix graph has a certain degree of similarity to all other learners. Figure 11 In .c and 11.f, the low-similarity matrix graph has only one connected component. Unlike the medium-similarity matrix graph, the connected component density of the low-similarity matrix graph is very low, and the edges in the graph mainly come from one or a few learners. This indicates that there are one or a few learners in the time slot with very low similarity to other learners. However, this does not mean that these learners have no learners with high similarity.

[0117] b) ITLGD: Based on previous experience, a modularity greater than 0.3 indicates that the partitioning is acceptable for community detection algorithms. Therefore, this invention sets an input modularity threshold for the algorithm. For time slot 1 of course 1, the algorithm output is... There are a total of 57 learning groups; There is one learning group; The number of students in the middle school group is 0. For course slot 2, 7... There are a total of 26 learning groups; There is one learning group; The number of students in the middle school group is 0.

[0118] It is evident that the algorithm output structures of the two cases are basically the same: multiple learning groups with high similarity and close internal connections, indicating that learners within these groups maintain a high degree of synchronization in terms of behavioral preferences and performance; one learning group with moderate similarity, indicating that learners within this group share some similarities, but none are particularly pronounced; and zero learning groups with low similarity, indicating that no learner is isolated within a time slot. This distribution of learning groups is consistent with the common understanding of learning group distribution in online learning classrooms, therefore, this invention considers this structure reasonable.

[0119] To gain a deeper understanding This invention identifies the most representative learners within a learning group by calculating eigenvector centrality, based on the shared prominent characteristics of learners in that group. It then extracts the top 5 learning resources by interaction frequency, the top 3 activity types by activity level, the top 2 learning behaviors, and key test scores for each learner, using these as basic group labeling information. For example, the learning groups with group number 1-0-0 in slot 1 of course 1 and group number 7-0-4 in slot 7 of course 2 are shown below. Figure 12 As shown in the diagram. The nodes marked in red are the representative nodes of the group.

[0120] The basic group label for 1-0-0 is This set of tags represents different granularities of information highlighting the learners' prominent tendencies, including learning behaviors related to preparation and knowledge acquisition; the types of activities involved in participating in course forums, browsing course pages, and downloading course resources; and accessing the 3rd forum resource, 2nd quiz resource, 0th platform homepage resource, 17th download resource, and 3rd course page resource. Furthermore, it indicates that the group's rating is 76 points and they tend to submit quizzes in advance within their time slot.

[0121] The basic group label for 7-0-4 is This set of tags indicates that learners in this group may exhibit different granularities of information, including learning behaviors related to learning preparation and interactive learning; the types of activities they engage in, such as browsing platform pages, participating in course forums, and searching for course resources; and visits to the 0th platform homepage resource, the 10th forum resource, the 192nd course page resource, the 0th forum resource, and the 36th course page resource. Furthermore, it indicates that the group's rating is 80 points and they tend to submit quizzes in advance within their time slot.

[0122] c) SMLS: To further enhance the interpretability of learning groups and comprehensively characterize their behavioral patterns and learning outcomes, this invention proposes a multi-dimensional tagging system based on basic group labeling information, combined with behavioral characteristics, academic performance, and homework habits. Similarly, taking groups 1-0-0 and 7-0-4 as examples, the comprehensive tag for 1-0-0 is Preparatory Knowledge Acquirer-Pass-Early, indicating that this group focuses on pre-class preparation and knowledge acquisition, achieves satisfactory academic results, and has excellent homework completion habits. The comprehensive tag for 7-0-4 is Collaborative Preparer-Excellent-Early, indicating that this group combines pre-class preparation with social interaction, achieves satisfactory academic results, and has excellent homework completion habits. Such comprehensive tags not only clearly express the group's multi-dimensional prominent characteristics but also provide personalized guidance for groups based on specific dimensional tags. For example, for groups whose performance-homework habit label is Fail-Late, education administrators can provide time management guidance; for groups whose behavior-performance label is Systematic Reviewer-Fail, administrators may need to intervene in a timely manner to provide these groups with the right learning guidance and improve the efficiency of output; for groups whose comprehensive label is Exploratory Collaborator-Excellent-Early, teachers can assign these groups open-ended group projects to encourage further cross-group collaboration.

[0123] Table 1

[0124] abbreviation meaning LITP Learner individual time slot profile sTP Sub-time slot portrait Demographic, De Sub-slot profile components, learner demographic information Materials, Ma Sub-slot profile components, learning resource preferences Categories, Ca Sub-slot profile components, activity type preferences Behaviors, Be Sub-slot profile components, learning behavior preferences AsTP-DV The degree of change in the sub-slot profile of learners between adjacent time slots OsTP-EP Overall Evolutionary Pattern of Learner Sub-Slot Profile Members Intra-TGP Group portraits in time slots ITLGD In-slot learning population partitioning algorithm SMLS Simple Multidimensional Labeling System

[0125] Table 2

[0126] Combination of behavioral characteristics Behavioral pattern tags meaning Acquisition, Preparation Preparatory knowledge acquirer Emphasis on pre-class preparation and resource research Interactive, Preparation Collaborative preparers Combining pre-class preparation with social interaction Consolidation, Interactive Interactive Consolidator Consolidate knowledge through collaborative activities and repeated practice. Acquisition, Interactive Exploratory Collaborators Actively explore resources and participate in interactions Consolidation, Preparation Systematic reviewers Systematic review combined with pre-class preparation Acquisition, Consolidation Highly efficient learners Focus on effective knowledge acquisition

[0127] Table 3

[0128] Learning behavior FFF2013J DDD2013J Preparation homepage, page, subpage, folder homepage, subpage Acquisition dataplus, glossary, oucontent, resource, url resource, url, glossary, oucontent Interactive ouwiki, oucollaborate, forumng, dualpane forumng, oucollaborate, ouwiki Consolidation repeat activity, quiz, questionnaire externalquiz

[0129] Table 4

[0130] Activity type meaning coding homepage Access the main interface of the learning platform H page Access the course interface P subpage Access the course sub-interface SP folder Access the course folder F dataplus Access the course extradata DPS glossary Access the glossary G oucontent Download the course resources CT resource Search the course resources RS url Access course URL link U ouwiki Query with Wikipedia WK oucollaborate Participate in collaborative exchange activities CB forumng Participate in the course topic forum F dualpane Access dual-pane interface DPE repeatactivity Repeat course activities / tasks RA quiz Complete curricular quizzes Q questionnaire Complete curricular questionnaires QN externalquiz Complete extracurricular quizzes EQ

Claims

1. A method for constructing and analyzing learner profiles based on time slots, characterized in that, Includes the following steps: Step 1: Data Preprocessing and Adaptive Time Slot Division; Obtain learners' raw log data from the online learning platform, including records of resource clicks, activity participation, quiz submissions, and demographic information; After cleaning and feature extraction of the data, an adaptive time slot division method is proposed: using the deadline of the weighted quiz, which has time uniqueness and is strictly synchronized with the teaching progress in the course, as the key node, the entire learning cycle is divided into N consecutive time slots; each time slot covers the complete learning cycle between two adjacent weighted quizzes, thus providing an ideal time scale for subsequent dynamic analysis; Step 2: Construct individual learner time slot profiles; For each time slot divided in Step 1, construct a sub-time slot profile for each learner; This profile is organized in a tree structure and contains three hierarchical components: learning resource preferences, and its activity level is calculated based on indicators such as the relative total number of clicks on resources and the number of interaction days; Activity type preference: Based on the interaction with the associated resources, calculate the preference ratio of different activity categories; Learning behavior preferences are calculated by aggregating and calculating the distribution of different learning behaviors based on activity type preferences; the profile of the first time slot additionally includes the learner's static demographic information; the complete individual time slot profile of the learner throughout the course is composed of the sub-time slot profiles of all its time slots in sequence; Step 3: Analyze the dynamic evolution of individual time slot profiles; to deeply analyze the dynamic changes in learners' preferences, two complementary analysis methods are proposed: adjacent time slot change degree analysis AsTP-DV and overall evolution pattern analysis OsTP-EP. The former calculates the difference in preference values ​​of each component of sTP (learning resource preference, Ma; activity type preference, Ca; learning behavior preference, Be) between adjacent time slots, and quantifies the local volatility and stability of learner preferences by calculating the entropy of the probability distribution of these differences; the latter treats the activity sequence of each component in N time slots as a whole, identifies three evolutionary sub-modes of "stable", "rising", and "falling" by defining "evolutionary segments" and setting thresholds; and summarizes six global evolutionary modes such as "high-activity stable type", "high-activity transition type", and "occasional participation type" by combining high / low activity states. Step 4: Construct a learner social network and calculate similarity within a time slot; For a single time slot, learners are abstracted as network nodes; By comprehensively calculating the similarity of interaction behaviors, performance similarity, and preference similarity among learners, and obtaining the total similarity by weighted summation, a weighted learner social network is constructed using this as the edge weight. Step 5: Implement the learning group division within the time slot; Based on the social network constructed in step 4, a learning group segmentation algorithm within time slots is proposed. First, learners are divided into three similarity groups: high, medium, and low, according to a similarity threshold. For the high and medium similarity groups, the graph structure is initialized and the Louvain community detection algorithm is applied to optimize modularity and perform multi-level community segmentation to identify highly cohesive secondary learning groups. Learners in the low similarity group are regarded as independent individual groups. Step 6: Generate group profiles and interpretable tags; First, by calculating the eigenvector centrality of nodes within the group, the most representative core learners are identified; then, the Top-N learning resources, activity types, learning behaviors, and key achievements of these representative nodes are extracted as basic group tags; finally, through a designed simple multi-dimensional tagging system, the semantic tags of three dimensions—behavioral patterns, academic performance, and homework habits—are integrated to generate comprehensive group tags in the form of "collaborative preparer - excellent - early submission," greatly enhancing the interpretability of the group profile; Step 7: Method Validation and Case Analysis; The above method is systematically validated on real public datasets; By showcasing the dynamic LITP of specific learners, the evolutionary entropy value map and overall pattern map of their sTP, as well as the group segmentation results and comprehensive labels of different time slots, the effectiveness and superiority of this method in capturing learner dynamic behavior, segmenting meaningful learning groups, and providing interpretable insights are demonstrated qualitatively and quantitatively.

2. The learner profile construction and analysis method based on time slots as described in claim 1, step 2 specifically involves: for N time slots... A certain time slot For example, the sub-slot profile of learner S is denoted as in, They represent time slots respectively. The start and end times; This indicates that the student's basic information consists of m features; Ma, Ca, and Be are respectively denoted as: This indicates that the time slot image contains n learning resources; , indicating that there are a total of p types of activities; Let q represent the number of learning behaviors; ultimately, the individual time-slot profile of learner S is denoted as q. ; To quantitatively describe learners' preferences for learning resources, activity types, and learning behaviors across different time slots, the activity level of each item is calculated to quantify learners' behavioral characteristics and learning habits; for a learning resource... Its activity level is calculated from four perspectives, using the formula. The calculation shows that, The percentage of the total number of clicks for the resource. The percentage of relative interaction days for resources The relative highest number of clicks for the resource and The longest consecutive number of days the resource has been used; the first three items are respectively derived from the formulas. Calculations show that It was derived from statistics; These are the corresponding weights, and . and All will be normalized; ； ； ； ； in, Representing resources The number of days that occur within a time slot; express Total number of clicks on day j; Indicates time slot The number of days; Each activity type has a percentage relative to other activity types in the same time slot; the percentage is determined by the activity category. The relative number of days of resource usage and the relative total number of clicks are determined as shown in the formula. As shown; ； in, Indicates that it belongs to the activity type A collection of learning resources Indicates learning resources In the time slot The number of days of interaction between China and the United States Indicates students' interest in learning resources The total number of clicks on day j of its interaction period; These are the corresponding weights, and their sum is 1; For learning behavior, the sum of behavioral preferences across time slots is 1; learning behavior From the formula Calculated; ； in, This indicates that it belongs to learning behavior. A collection of activity types; After the above steps, a sub-slot profile of the learner is constructed; by repeating this process N times, a complete individual slot profile of the learner is finally obtained.

3. The learner profile construction and analysis method based on time slots as described in claim 1, step 3 specifically comprises: AsTP-DV focuses on the differences in sub-slot profiles between adjacent time slots. It assesses the stability of learners' preferences across different time slots by quantifying the probability distribution of these differences and calculating entropy values. First, it extracts learners' sub-slot profile member preferences from time slot 0 to time slot N from their individual time slot profiles. Time slot 0 is defined as the period from course release to the start of the course. During this stage, the platform still records learners' interaction data with the course; therefore, the data from time slot 0 is considered as learners' pre-learning behavior. A learner's sub-slot profile consists of three parts: adjacent time slots... The difference in sub-slot profiles is defined as follows: ,in These are the differences in preferences for learning resources, activity types, and learning behaviors across adjacent time slots; among the activity type preferences, for N+1 time slots... The activity type set is In time slots In the context, the set of activity types is ,Then, ,in express Time slot Activity level minus Time slot The activity level; a total of N+1 time slots have indivual Then, for each pair of adjacent time slots in the N+1 time slots, the probability distribution of the difference between each pair of time slots is first calculated, and then the stability of these distributions is described by the entropy value; the entropy value can reflect the degree of fluctuation of the changes between time slots; the larger the entropy value, the more significant the change of the learner's preference between time slots, and the smaller the entropy value, the more stable the change of the learner's preference between time slots. OsTP-EP quantifies the analysis of learners' inter-slot evolution patterns as the analysis of the evolution of a series of numerical sequences; the numerical sequence refers to the activity sequence composed of various preferences of learners across N time slots; in learners' activity type preferences, different evolutionary segments are defined based on three evolutionary sub-patterns. ,in IV and h represent the initial value of the evolutionary segment and the difference between the two nodes, respectively; Define threshold =0.01 to distinguish evolutionary subpatterns: If a certain EF in Then the evolutionary segment ;like ,but ;like ,but . For a specific type of activity ,in Indicates a specific activity type In the time slot Activity level; Represented by a sequence of N-1 evolutionary segments, The proportion is given by the formula Calculated; ； Dynamic thresholds are set for different activity types, categorizing activity levels into high and low activity levels. ,but This time slot represents a high-activity type of activity. Conversely, low-activity activity types. , The proportion is given by the formula Calculated; 。 4. The learner profile construction and analysis method based on time slots as described in claim 1, step 4 specifically comprises: comprehensively considering the similarity of learners' preferences, test scores, and interactive behaviors; the total similarity between learners x and y in time slot i is defined as the weighted sum of the three similarities: interactive behavior similarity. Similarity of scores Similarity to preferences , as in the formula As shown, where . ； Similarity of interactive behaviors The calculation is performed by analyzing learners' interaction behavior with learning resources within time slots; firstly, a... The matrix records the actions of all learners Ls in time slot i within that time slot. The number of clicks on the learning resource Let represent the set of learning resources that learner j has interacted with in time slot i, and let the value of the matrix be the number of times the learner accesses the learning resources; then, use cosine similarity to calculate the similarity of the interaction behavior between any two learners. similarity of scores This is calculated by analyzing the learners' test score vectors; each learner maintains a score vector of length 5. ； in, numAsmts represents the number of tests; TMAAsmtScore represents the score of the weighted test; TMAAsmtSubDate represents the submission date of the weighted test; avgCMAAsmtScore represents the average score of the unweighted test; avgCMASubDate represents the average submission date of the unweighted test; the similarity of scores among learners is also calculated using cosine similarity. Preference similarity Reflecting learners' interests and preferences in learning resources and activity types; each learner maintains a database of length [length missing]. The preference vector, where This represents the number of learning resources that learner j interacted with within time slot i. The vector represents the number of activity types that learner j participates in within time slot i; the values ​​of the vector represent the learner's activity level towards learning resources and activity types; the preference similarity between learners is calculated using cosine similarity.

5. The learner profile construction and analysis method based on time slots as described in claim 1, step 5 specifically comprises: First, initializing an empty graph for each similarity group, and filling it in descending order of similarity between learners; for high similarity groups, learners are added to the graph as nodes and similarity values ​​as edges; since the graph may contain multiple connected components, a modularity threshold is set; for connected components with a modularity less than the threshold, the partitioning is cancelled, and the entire connected component is treated as a group and added to the learning group; for connected components with a modularity greater than or equal to the threshold, each community within it is divided into an independent group; Second, when filling the middle similarity groups, ensuring that learners already belonging to the high similarity group are not added repeatedly, and using the same partitioning method as the high similarity group for group construction; Finally, for low similarity groups, ensuring that learners from the first two groups are not included, each learner in the low similarity group is treated as an independent group; By calculating the eigenvector centrality, the most representative learners in the group are identified, and based on these learners, the top-5 learning resources with the most interactions, the top-3 activity types with the highest activity levels, the top-2 learning behaviors, and the weighted test scores are extracted as basic group label information.

6. The method for constructing and analyzing learner profiles based on time slots as described in claim 1, characterized in that, Step 6 specifically involves proposing a multi-dimensional tagging system based on basic group label information, combined with behavioral characteristics, academic performance, and homework habits; Based on the following three core dimensions: behavioral patterns, academic performance, and homework submission habits, an interpretable semantic tag is ultimately formed to support the design of teaching strategies and personalized interventions. There is a hierarchical relationship between learning behaviors, activity types, and learning resources, with granularity ranging from coarse to fine. Considering the generalizability of this tagging system across different courses, the coarsest-grained learning behavior is taken as the behavioral feature; The online learning process includes the learning phase, the knowledge acquisition phase, the interactive reflection phase, and the learning consolidation phase. The learning phase is the preparatory process for learners to formally begin online learning; The knowledge acquisition stage is the most important part of the online learning process, and it is also the process by which learners initially acquire knowledge. The interactive reflection stage is a process in which learners interact with teachers and peers and engage in self-reflection during the interaction. The learning consolidation stage is the process by which learners consolidate and internalize knowledge.

Images