Pattern Recognition and Intelligent Analysis of Large-Scale Student Behavior Data in the Perspective of Educational Technology

With the continuous development and application of science and technology, the position and role of educational technology in the field of education and other related fields are becoming more and more prominent. Its role is to support the realization of modern educational thinking, enrich the content of curriculum and teaching, provide diversified methods and optimize the process of teaching and learning [1-2]. In practical application, how to identify and analyze students’ behavioral data in the field of educational technology plays an important role in today’s education system, education methods and education management.

First, pattern recognition and intelligent analysis of students’ behavioral data can help to deeply explore students’ learning characteristics and behavioral patterns, and provide a scientific basis for educational practice. Collecting students’ behavioral data through school management systems, online education platforms, questionnaire surveys and interviews, intelligent devices and applications, etc., we can understand students’ attitudes, points of interest, difficulties, etc., which can help educators to better formulate teaching strategies and improve educational effects [3-5]. Among them, intelligent analysis can mine the interrelationships between different attributes of students’ behavioral data through association rules, such as the correlation between the frequency of students’ participation in extracurricular activities and the level of academic performance [6-8]; it can also be used to classify students into different types through clustering analysis, such as diligent, lazy, potential, etc., so that we can provide targeted educational measures for different types of students and select the most suitable employees for enterprises [9-11]; predictive analytics can also be used to predict students’ academic performance, probability of further education, chances of employment, etc [12-15]. Second, pattern recognition and intelligent analysis of student behavior data are conducive to the realization of personalized education. Through the intelligent analysis of students’ behavioral data, we can understand the characteristics of students’ learning ability, habits, personality, interests, etc., and provide students with personalized learning and guidance services to meet their different needs [16-17]. Thirdly, pattern recognition and intelligent analysis of student behavior data can help improve the refinement and intelligence of education management. Through the collection and analysis of students’ behavioral data, the problems of students in learning can be found and solved in time, and at the same time, it can provide decision-making support and scientific basis for educational administrators [18-19].

It can be seen that student behavioral data is widely used, and the pattern recognition and intelligent analysis of them are extremely meaningful in the field of education.

This paper proposes an intelligent analysis method combining the K-Means algorithm and lag sequence analysis to analyze students’ online learning large-scale behavioral data. Data mining technology is used to collect online behavioral data from students, extract the actual and effective behavioral features, and use K-Means clustering to classify different learning groups based on the above behavioral features. Then, using the lag sequence analysis method, the sequence of students’ learning behaviors is analyzed to explore the subtle differences in their behaviors. In the comparison of learning behavior differences and sequence analysis, methods such as literature collation and case study analysis are used to comprehensively and systematically analyze students’ online learning behaviors.

2

Methods for identifying and analyzing student behavioral data

2.1

Student Behavioral Data Recognition Technology

2.1.1

Data mining

With the continuous development of the Internet, online learning platforms have accumulated more and more behavioral data with high educational value, and the process of extracting potentially regular and useful information and knowledge from these data that people did not know before is data mining [20], and the main process is shown in Figure 1.

Data mining is categorized into the following five stages, and the following are the specifics of each stage. 1)

Define the goal. Confirm what problems are to be solved using data mining, clear objectives are the basis for successful data mining.

2)

Data extraction. Collect data sets related to the data mining objectives, but not all data are available for data mining, the need to eliminate invalid and redundant data, from which to select effective and suitable data.

3)

Data preprocessing. Due to the inconsistencies and missing values of the obtained data, which cannot be directly used for analysis, preprocessing is required before data analysis, and preprocessing includes steps such as data cleaning, data integration, and data specification, first of all, data review and verification, check the consistency of data, and solve the problems of missing and invalid values, so as to eliminate the impact of “dirty data” on the data extraction process.

4)

Model discovery. Select the appropriate algorithm to mine the data that has been processed and transformed, train the model, construct the corresponding data model, and evaluate and interpret the model based on certain assessment indicators.

5)

Knowledge representation. After completing the above steps, the useful information and knowledge required by the target will be obtained in the end.

2.1.2

Data generalization

Descriptive mining in data mining mainly focuses on describing the data concisely and providing relevant properties and features of the data. Data generalization, on the other hand, is one of the methods of descriptive mining, whereby data is processed to obtain concise information about the data description. 1)

Data Focusing. By selecting the attributes and dimensions that are relevant to the analysis and choosing the appropriate dataset, the extraction process can be made more efficient and the results more meaningful. In descriptive mining process, generalization efficiency is improved by filtering irrelevant or weakly relevant attributes and thus.

2)

Attribute Removal. If an attribute of an initial working relationship contains a large number of different values, the attribute should be deleted in the following two cases: a) No generalization operator is found on the attribute, it cannot be generalized, and its retention would be contrary to the goal of generating concise rules. b) Higher-level concepts of the attribute can be represented by other attributes, and its deletion is equivalent to its generalization operation.

3)

Attribute Generalization. Is an important way of dealing with attributes that have different values and the process can be transformed using the generalization operator. If attributes are generalized too deeply, it can lead to overgeneralization.

4)

Attribute generalization threshold control. The generalization threshold of an attribute is the maximum number of distinct values allowed for that attribute, beyond which it needs to be further removed or generalized. One threshold can be set for each attribute or the same threshold can be set for all attributes.

5)

Generalized Relationship Threshold Control. If the number of different tuples in the generalized relationship exceeds the threshold, generalization should continue. Otherwise no more generalization. The data mining system can also provide this threshold (default value range 10~30), or expert control. In practice, you can choose which control method to use initially based on the user’s needs, and then make adjustments as required.

6)

Representation of generalization. Generalization is a way of dealing with raw data that can be used to derive generalization relations. Typically, generalization relationships can be presented directly to the user in the form of a final conceptual description, or generalization relationships can be presented in the form of cross-tabulations, pie charts, bar charts, bar graphs, curves, or quantitative feature rules to provide a more intuitive or abstract description of the generalization results.

The most common means of data mining is correlation analysis, which is mainly used to explore the degree of correlation and mutual influence between multiple data variables and to interpret the correlation analysis.

Let the distribution function of the binary aggregate (X, Y) be the variances of F(X, Y) are var(X) and var(Y), respectively, the overall covariance is cov(X, Y), and the overall correlation coefficient is defined as: 1 $ρ_{X Y} = \frac{cov (X Y)}{\sqrt{var (X)} \cdot \sqrt{var (Y)}}$

Let $(X_{I}, Y_{I}), (X_{2}, Y_{2}), \dots, (X_{n}, Y_{n})$ be an independent sample selected from each of the binary totals (X, Y), and the sample correlations can be computed: 2 $r_{X Y} = \frac{S_{x γ}^{2}}{\sqrt{S_{x}^{2}} \cdot \sqrt{S_{Y}^{2}}}$

where $S_{X}^{2}$ and $S_{Y}^{2}$ are the variances of samples X and Y, respectively, and $S_{X Y}^{2}$ is the covariance of sample XY. In the usual case, r_XY computed from the samples is not zero, even when the random variables X, Y are independent. Therefore when ρ_XY = 0, it is not practically meaningful to use r_XY to measure the correlation of X, Y, so hypothesis testing is needed: 3 $\begin{matrix} H_{0} : ρ_{X Y} = 0 \\ H_{1} : ρ_{X Y} \neq 0 \end{matrix}$

It can be shown that the statistic when (X, Y) is a binary normal totality and when H₀ is true: 4 $t = \frac{r_{X Y} \sqrt{n - 2}}{1 - r_{X Y}^{2}}$

Obeys a t distribution with n − 2 degrees of freedom.

Using the property that the statistic t obeys a t distribution with n − 2 degrees of freedom, the data X, Y can be tested for correlation. Since the correlation coefficient r_XY is known as the Pearson (Pearson) correlation coefficient, the test is also known as the Pearson correlation test.

2.1.3

Extraction of online learning behavioral features

According to the theory of behavioral science, learners use online learning platforms to fulfill their own learning needs, and behavioral science focuses on observable and measurable episodic behavioral activities. In the process of online learning, learners’ operations can be observed and quantified, so the analysis of online learning behavior takes online operations as a breakthrough, and is studied by observing, describing and refining learners’ learning operations.

It has been discovered that effective learning behaviors are usually classified into four types. In this study, with reference to the actual effective behaviors produced by learners when learning online, and with reference to the summary of related literature, a suitable behavioral analysis framework is constructed to classify learners’ online learning behaviors into four categories: time investment, learning effectiveness, learning interaction and learning motivation. Eleven specific behavioral indicators were extracted.

According to educational psychologists’ categorization of learning outcomes, learning through interaction fits both the category of verbal learning and attitude, i.e., learners’ feedback to themselves and others through verbalized knowledge and on acquired knowledge. The eleven behavioral characteristics listed basically cover the dynamic behaviors of existing e-learning platforms in their entirety (the point in time at which each behavior occurs is not addressed in this study), and the remaining characteristics that are not included can be considered to be deformations, extensions, or amalgamations of the listed characteristics.

2.2

Methods for Intelligent Analysis of Student Behavior Data

2.2.1

Cluster analysis

Clustering is an algorithm involving data grouping in data mining, which is an unsupervised classification method that reveals the intrinsic properties and patterns of data through the learning of unlabeled training samples to prepare for subsequent data analysis. It is mainly used to measure the similarity between different data sources, categorize data sources into different clusters according to the principle of grouping things into clusters, and find key points and influencing factors. Cluster analysis, also known as cluster analysis, is a statistical analysis method to study the problems of samples or indicators. It aims to divide the set of samples with unknown markers into groups according to certain guidelines, with the aim of achieving high similarity within groups and low similarity between groups.

In the identification of unknown learning patterns for online learners, it is usually necessary to find the association from a bunch of disordered data and find the data with similarities, which is an important application of cluster analysis in the process of e-learning, and has an important reference value for discovering learning patterns and adjusting learning strategies. 1)

Hierarchical clustering

Hierarchical clustering [21] involves creating clusters according to the stratification of the data to form a tree with clusters as nodes, i.e., a clustering graph. It has the advantage of being able to draw a tree diagram which helps in interpreting the results of clustering using a meaningful taxonomy, another advantage of hierarchical clustering is that there is no need to specify the number of clusters in advance. Hierarchical decomposition is further classified into cohesive hierarchical clustering and split hierarchical clustering based on whether it is bottom-up (merging) or top-down (decomposition). Cohesive hierarchical clustering starts by treating each object as an independent cluster, and gradually merges the clusters according to certain conditions until all similar objects are in one cluster or until a certain termination condition is met. Split hierarchical clustering is the opposite, which places all objects in the same cluster at the beginning and then divides a cluster into finer clusters according to conditions until a termination condition is reached. Common inter-cluster metrics include minimum distance (single linkage method), maximum distance (full linkage method), and average distance (average linkage method).In this study, cohesive hierarchical clustering is mainly used, in which each sample is viewed as a separate cluster in the analysis, and then the closest pair of clusters are merged repeatedly until the termination clustering condition is reached. The study used the Ward method for hierarchical clustering. Euclidean distance was used to measure intervals and generate a genealogical map to categorize the learners into groups of online learners.

The formula for calculating the interval between data points of different categories using Euclidean distance is given below: 5 $D = \sqrt{{(x_{1} - y_{1})}^{2} + {(x_{2} - y_{2})}^{2}}$

2)

K-Means Clustering

K-Means [22], also known as k-mean clustering algorithm, belongs to unsupervised learning. In unsupervised learning, the labeling information of the training samples is unknown, and the algorithm reveals the properties and patterns embedded in the data by learning from unlabeled samples. The main purpose of clustering algorithm is to classify the similar data in the dataset into the same cluster based on the data characteristics.K-Means algorithm is a very widely used clustering algorithm. The idea of its algorithm is also very simple. It usually uses the Euclidean distance as the evaluation index of similarity, which means that the closer the clustering of two objects is, the greater the similarity. At the same time, k-mean clustering needs to determine the constant k in advance, which is also the final number of clustering categories, and its selection has a greater impact on the clustering results. The initial point was randomly selected as the center of mass, and by calculating the similarity between each sample and the center of mass (using the Euclidean distance method), the sample points were classified into the most similar class, then the center of mass was recalculated, and the process was repeated until the center of mass was no longer changed, which ultimately made it possible to determine the class to which each sample belonged and the center of mass of each class. In addition, the K-Means algorithm tends to converge slowly in large-scale datasets due to the repetitive calculation of similarity between each sample and the center of mass.

The K-Means algorithm is an optimization algorithm that is updated iteratively and needs to ultimately result in the smallest possible mean square error, as shown below: 6 $\min \sum_{i = 0}^{k} \sum_{x_{j} \in c_{i}} ({‖ x_{j} - u_{i} ‖}^{2})$

where x_j is the jnd data object; u_i is the ith clustering center; and k is the number of clusters divided.

2.2.2

Lagged Sequence Analysis Methods

Lagged Sequence Analysis (LSA) is mainly used to test the probability of one behavior occurring after another in people and whether it is statistically significant or not. Currently, LSA has been applied to customer behavioral preference analysis in the field of e-commerce, patient behavioral analysis and treatment in the medical field, and player game behavior analysis in the gaming field.

LSA is used in learning behavior analysis, which not only provides a new analysis method for verifying the effectiveness of new technological tools, models, methods and strategies, but also provides a new perspective for explaining the changes in students’ learning performance in the field of technology-enhanced learning research, and at the same time provides a new basis for teachers to carry out more targeted and personalized teaching and counseling. Lagged Sequence Analysis opens up new paths for learning analysis, making online learning more effectively meaningful and clearer in its understanding of the learner’s learning process and his or her own behavioral motivations. But in the process of utilizing LSA, it needs to be used in conjunction with other research methods in order to carry out the research. LSA cannot exist independently and lose its significance. Therefore, in the research process, LSA is generally used in conjunction with other methods (such as questionnaires, tests, interviews, etc.) in order to interpret the learning process from multiple perspectives and analyze the results of the study. In addition, the application of LSA also needs to consider specific research issues, such as there are some learning activities with a specific process, there is no need to use the LSA analysis method, because the nature of this learning activity is prescriptive, in which case, LSA loses his unique meaning.

2.2.3

LSA-based learning behavior analysis

In this study, the behavior of students when they clicked on each option in the module was used as the behavior that occurred and coded based on the main behavioral data collected from students in each module in the mini-classroom platform, and the coding results are shown in Table 1.

Table 1.

Student learning behavior framework

Behavior classification	Behavior indicator	Behavior coding
Time input	Mission completion	A1
Time input	Test duration	A2
Learning results	Test score	B1
	Job score	B2
	Test effectiveness (accuracy/cost time)	B3
	Test results	B4
Learning interaction	Post number	C1
	Reply number	C2
	Discussion score	C3
Learning initiative	The first login delay	D1
Learning initiative	Task delay days	D2

After coding the behaviors it is convenient to count the frequency of occurrence of each behavioral sequence, then count the transition probabilities between each behavioral sequence, calculate the z-value, and filter the significant behavioral sequences based on the results of calculating the z-value. (Behavioral sequences with a z-value greater than or equal to 1.96 are significant behavioral sequences) 7 $Z_{j} = \frac{x_{i j} - E_{i j}}{\sqrt{E_{i j} (1 - P_{i}) (1 - P_{j})}}$ 8 $E_{i j} = \frac{\sum_{i} x_{i j} \sum_{j} x_{i j}}{\sum_{i} \sum_{j} x_{i j}}$ 9 $P_{i} = \frac{\sum_{i} x_{i j}}{\sum_{i j} x_{i j}}$ 10 $P_{j} = \frac{\sum_{j} x_{i j}}{\sum_{i} \sum_{j} x_{i j}}$

x_ij: Frequency of switching from behavior i to behavior j.

p_i: Probability of switching from behavior i to other behaviors.

p_j: Probability of switching from other behaviors to behavior j.

E_ij: The expected value of switching from behavior i to behavior j.

Finally, the value of z is obtained to determine the significance of behavior i. In this paper, the significance level is set to 0.05 and one-tailed test is used, then when the value of z is greater than 1.96, the transfer of the behavior can be regarded as significant. Finally, we can draw a behavioral transfer matrix, visualize the z value to present the analysis results, and we can also draw the significant behavioral transfer into a state behavioral transfer chart to show it more intuitively.

3

Comparison of Behavioral Differences in Different Learning Groups

Through cluster analysis, combined with the indicators of learners’ online learning behaviors, learners are clustered into four categories, and the results show that there are obvious differences between each category of learning groups. In order to analyze the specific performance of behavioral differences, the quantitative mean value of online learning behavior of each category is first compared, and the comparison results are shown in Table 2.

Table 2.

The mean of learning behavior online

	Learning results				Time input		Learning interaction			Learning initiative
	B1	B2	B3	B4	A1	A2	C1	C2	C3	D1	D2
1	57.62	100	92.13	0.012%	5751.11	12002.76	0.81	1.41	2.02	12.53	5.71
2	50.11	94	75	0.062%	1051.24	9057.87	0.42	1.28	0.03	13.77	6.83
3	55.24	97.11	88.56	0.034%	2480.63	10925.25	1.34	13.08	4.57	11.54	4.82
4	68.02	93.22	90.16	0.195%	3207.56	13930.46	1.01	8.22	2.92	9.06	2.26

According to the mean values of catechism learning performance, learning hours, interaction times and learning delay days of each category of learning groups, the following definitions are made for each category of learners: 1)

The learning group of category 1 performs well in general, achieves excellent ordinary grades, and is comparable to category 4 in terms of learning interaction performance, but the learning group’s learning procrastination is higher, so the learning group is defined as “average learners”.

2)

The overall performance of the learning group in category 2 is poor, and the observation table shows that the test effectiveness of this group is higher, but the specific analysis is due to the lower average test length of this group. Moreover, the procrastination of this group is high, so this group is defined as “negative learners”.

3)

The overall performance of Category 3 learners is comparable to that of Category 4, with good results and more active completion of tasks. However, in terms of learning interactions, this group achieved the highest results in discussions and actively replied to posts in the forum, so this group was defined as “interactive learners”.

4)

Learners in category 4 achieved excellent overall results, had the highest test efficacy among the four categories, participated more actively in topic interactions, and had low procrastination in completing tasks, thus defining this group as “active learners”.

The study chooses two factors that are more important to the clustering: the total time to complete the task and the number of replies to the scatterplot, and the results are shown in Figure 2, which shows that there is a significant difference between the various types of learning groups in these two factors. The analysis found that “negative learners” performed poorly in both the total time to complete the task and the number of replies, and “interactive learners” had the highest number of replies among the four types of learning groups.

3.1

Comparing learning motivation

In order to analyze the differences between each type of learning behaviors of each type of learning groups, this study uses ANOVA to explore the differences between each type of learning behaviors using the LSD method.

Table 3 shows the results of the ANOVA of the learning motivation behaviors of different learning groups, and it can be seen that in the dimension of learning motivation, except for “general learners” and “interactive learners”, there are significant differences between the two behavioral indicators of learning motivation for the rest of the learning groups. Differences. In terms of the number of days delayed in tasks, the difference between “interactive learners” and “general learners” and “negative learners” is not significant.

Table 3.

Study the learning behavior of different learning groups

Significance
Learning behavior	Class number	2	3	4
The first login delay	1	0.025	0.198	0.000
	2		0.006	0.019
	3			0.008
Task delay days	1	0.000	0.048	0.000
	2		0.039	0.000
	3			0.000

The statistics of the time interval between the first login of different learning groups are shown in Figure 3. Compared with the other three groups of learners, the overall log-in interval of “active learners” is smaller, and the number of first-time log-in intervals of 0 days reaches 25, indicating that this group of learners actively logs in and participates in learning.

3.2

Comparing time commitment

First of all, we analyze the differences in learning behaviors of different learning groups in the category of time investment, and the results of ANOVA are shown in Table 4, which reveals that there is a significant difference between all types of learning groups. On the index of test duration, the difference between “active learners” and “interactive learners” is smaller than that of other groups, with a significance of 0.028.

Table 4.

Time input learning behavior different learning group LSD test

Significance
Learning behavior	Class number	2	3	4
Mission completion	1	0.000	0.000	0.000
	2		0.000	0.000
	3			0.000
Test duration	1	0.000	0.000	0.000
	2		0.002	0.000
	3			0.028

Figure 4 shows the statistics of the total time for different learning groups to complete the task, and it is found that compared with the other three types of learning groups, “active learners” have the longest online learning time, and the number of students with 14,211 hours of learning is 22, which indicates that they have the highest commitment to learning. On the contrary, “negative learners” consume the shortest time, although the test effectiveness of this group is relatively high, but due to the low consumption of time in the test, which confirms that this group is less motivated to learn.

3.3

Comparing learning outcomes

Learners’ learning effectiveness often becomes the only criterion by which instructors judge their effectiveness. The study first explored the differences between different learning groups on the dimension of learning effectiveness, and the results of the ANOVA are shown in Table 5. Among the indicators of learning effectiveness, the differences between categories were relatively small, and there was only a difference between “active learners” and “average learners” and “passive learners” in the scores of practice questions. There were significant differences between the “active learners” and the other three groups of learners, while there were no significant differences between the other three groups of learners. In terms of test scores, there were significant differences between the “passive learners” and the other three learning groups, while there were no significant differences between the other three learning groups. There were significant differences in test effectiveness among different learning groups. It follows that it is not reasonable to use learning effectiveness as the only criterion for assessing learners.

Table 5.

Learning behavior of different learning groups

Significance
Learning behavior	Class number	2	3	4
Exercise score	1	0.388	0.789	0.018
	2		0.638	0.026
	3			0.098
Normal grade	1	0.096	0.325	0.000
	2		0.569	0.000
	3			0.000
Test results	1	0.000	0.322	0.083
	2		0.000	0.000
	3			0.621
Test efficiency	1	0.008	0.045	0.034
	2		0.005	0.004
	3			0.045

When comparing the differences in test scores among the various categories of learning groups, the differences in test scores are shown in Figure 5, where the peak of the test scores of “average learners” is centered to the right of 88 points. The test scores of “negative learners” are significantly different from those of the other three groups, with a peak score of 73, which is located on the leftmost side of the four groups, indicating that this type of learning group has not achieved the desired learning results. By clustering the learning groups and analyzing the differences between different learning groups, it is easy for teachers to detect abnormal behavior in time and take targeted measures.

3.4

Comparative learning interactions

The number of text discussions reflects the learning status of learners and their activity during the course, and the study conducted ANOVA on the behavioral indicators of learning interactions of different learning groups, and the results are shown in Table 6. The results are shown in Table 6. It can be seen that there are significant differences between “interactive learners” and the other three groups in the three behavioral indicators of learning interaction, especially in the frequency of replying to posts, which is the most significant difference, confirming the analysis above.

Table 6.

Learning behavior of different learning groups

Significance
Learning behavior	Class number	2	3	4
Posting	1	0.015	0.000	0.000
	2		0.000	0.000
	3			0.007
Reply number	1	0.638	0.000	0.000
	2		0.000	0.000
	3			0.000
Discussion score	1	0.000	0.000	0.978
	2		0.000	0.000
	3			0.025

Specifically analyzing the differences in the frequency of each group of learners as shown in Figure 6, the peak number of “interactive learners” is 13 times, which is located in the most right-hand side of the four categories, indicating that this group of learners has the highest frequency of interaction. Combined with the above analysis, it can be seen that although this learning group actively participates in the interaction, but the achievement is not the most ideal, and at the same time, in connection with the interactive performance of “active learners”, it can be seen that if the more evenly the distribution of online learning behaviors of the learners, and do not focus on favoring a certain learning behaviors, it will be easier to achieve the desired goals.

4

Sequential analysis of classroom learning behaviors of different learning groups

Students’ behavioral sequence transitions occurring on the learning platform are diverse, with more significant sequences in the six learning behaviors of A1 (total hours of task completion), A2 (hours of testing), B4 (test scores), C1 (number of posts), C2 (number of replies), and D1 (number of days of delay in the first login). The conversion of students’ behavioral sequences can be used to discover learners’ potential learning behavior habits, patterns, and preferences.

4.1

Sequential Analysis of Learning Behavior of General Learners

The behavioral sequence of learners with average grades is shown in Table 7, which reveals that they perform better in exams and invest more in time investment, and their exam results are located at a more superior level, with a higher score of Z-Score=35.56 for the exam performance sequence, and the number of posts and replies to posts is also higher, which is comparable to that of active learners. In conclusion, general learners also invested more in learning interactive behaviors and favored the investment of knowledge time cost as much as active learning students, and general learners had more superior academic performance compared to active learners.

Table 7.

General learner behavior sequence residual value

Given	A1	A2	B4	C1	C2	D1
A1	22.81	-6.89	0.89	-4.98	-2.59	-5.61
A2	-3.58	28.96	4.56	2.65	4.28	4.12
B4	-1.56	-3.65	35.65	-2.59	-2.36	-2.33
C1	-4.85	-5.98	-4.71	24.56	-4.31	-3.65
C2	-5.69	-5.77	-1.25	-1.56	22.47	-4.25
D1	-8.69	-6.78	-11.36	3.69	2.65	14.65

4.2

Sequential Analysis of Learning Behavior of Negative Learners

The learning process of the negative learners is shown in Table 8, compared with the active learners and general learners, the negative learners’ learning motivation is not high, the total length of time for task completion is lower, and the test scores are lower, in terms of learning interaction, the number of replies and the number of posts are much lower than those of the other several categories of learners, and the Z-Score on the order of the test scores is 1.23. In a word, the negative learners s learning mode is more homogeneous, their investment in learning motivation behavior, time investment behavior, learning effectiveness behavior, and learning interaction behavior are all lower, and they are more lazy in learning.

Table 8.

Passive learner behavior sequence residual value

Given	A1	A2	B4	C1	C2	D1
A1	2.46	-5.9	2.66	-4.43	-2.51	-5.79
A2	-3.5	2.96	0.71	1.96	6.68	8.16
B4	-0.95	-3.07	1.23	-4.01	-3.06	-3.73
C1	-5.13	-6.96	-4.37	1.21	-4.91	-2.84
C2	-5.97	-5.38	-1.69	-1.54	2.81	-4.12
D1	-9.47	-5.7	-10.34	3.01	3.44	11.82

4.3

Sequential Analysis of Learning Behaviors of Interactive Learners

The learning process of interactive learners is shown in Table 9, and interactive learners are defined as students who prefer collaborative learning, who know their tasks in activities, accept feedback and comments from others, are able to achieve good grades, and are more active in completing tasks. However, in terms of learning interactions, this group of learners achieved the highest discussion scores and actively replied to posts on the forum.

Table 9.

Interactive learners behavior sequence frequency

Given	A1	A2	B4	C1	C2	D1
A1	9.01	-6.33	3.86	-2.83	-1.65	-5.42
A2	-4.05	7.2	-1.05	2.65	6.13	8.65
B4	0.19	-3.86	8.95	-3.58	-2.19	-4.79
C1	-3.91	-10.09	-5.85	39.56	-5.65	-3.18
C2	-4.46	-4.2	1.29	-1.75	38.97	-4.11
D1	-8.48	-10.37	-10.36	3.62	4.03	6.46

Interactive learners have a high Z-Score of 38.97 and 39.56 for the number of replies and posting sequence in learning behavior, which shows that interactive learners and teachers spend a large percentage of their time interacting with each other. Their grades, learning motivation, and engagement are at a medium level.

4.4

Sequential Analysis of Learning Behaviors of Active Learners

The behavioral sequence of active learners is shown in Table 10, students have the highest total hours of task completion in the learning process, the Z-Score of this behavioral sequence is 41.69, their behavioral pattern is more inclined to invest more time and their motivation to learn is higher. The number of days delayed for the first login was lower, and the Z-Score for this behavioral sequence was 2.13. In conclusion, active learners invested in interactive learning behaviors second only to interactive learning, were highly motivated to learn, had average academic performance, and invested the most time and were the most motivated to learn.

Table 10.

Active learner behavior sequence residual value

Given	A1	A2	B4	C1	C2	D1
A1	41.69	-8.02	5.73	-3.35	1.01	-3.16
A2	-4.91	37.2	-2.41	4.1	7.2	8.79
B4	-1.1	-4.42	18.95	-2.23	-3.34	-5.88
C1	-5.55	-7.08	-5.5	21.56	-5.05	-2.39
C2	-3.56	-5.68	2.35	-1.97	22.97	-1.86
D1	-10.28	-10.62	-9.81	5.43	1.59	2.13

5

Conclusion

In this paper, we use cluster analysis and lagged sequence analysis to identify patterns and intelligently analyze large-scale student behavior data, supported by educational technology, to provide guiding directions for classroom management.

Learners were categorized into four categories: average, negative, interactive, and positive through cluster analysis. These four categories of learners differed significantly in two factors: the total time spent on the task and the number of replies. “Average learners” performed better in terms of learning effectiveness, with a score center of more than 88 points. Passive learners performed poorly in terms of total time to complete tasks and the number of replies. Both “interactive learners” and indicators related to learning interaction performed well, with the highest number of replies among the four learning groups. “Active learners” perform best in terms of time commitment and motivation to learn. It suggests that the more evenly distributed learner behaviors are, the easier it is to achieve the desired learning effect, and the learning effect is not related to a single behavior.

The lag sequence analysis of students’ learning behaviors reveals that positive learners and interactive learners have more interactive behaviors, time investment, and learning motivation behaviors, while negative learners have a single learning mode and lower investment in the four types of learning behaviors. According to the situation reflected by these behavioral sequences, the teacher can further improve their teaching and put forward measures to enhance students’ learning motivation, interactive behaviors, and time investment. According to these behavioral sequences, teachers can further improve their teaching by proposing measures to enhance students’ learning motivation, interactive behavior and time investment, and at the same time, make the frequency of the above behaviors evenly distributed, so as to ensure that students can obtain good learning results.

Sprache:: Englisch

Zeitrahmen der Veröffentlichung:: 1 Hefte pro Jahr
Fachgebiete der Zeitschrift:: Biologie, Biologie, andere, Mathematik, Angewandte Mathematik, Mathematik, Allgemeines, Physik, Physik, andere

Zeitschrift RSS Feed

Pattern Recognition and Intelligent Analysis of Large-Scale Student Behavior Data in the Perspective of Educational Technology

Fangrui Li

Online veröffentlicht: 21. März 2025

Eingereicht: 06. Nov. 2024

Akzeptiert: 10. Feb. 2025

DOI: https://doi.org/10.2478/amns-2025-0568

SchlüsselwörterK-Means, Lagged Sequence Analysis, Behavioral Feature Extraction, Learning Behavior Sequence

© 2025 Fangrui Li, published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Schlüsselwörter
K-Means, Lagged Sequence Analysis, Behavioral Feature Extraction, Learning Behavior Sequence