Higher Education in China during the Pandemic: Analyzing Online Self-Learning Motivation Using Bayesian Networks

Abstract

The COVID-19 pandemic has led to an unprecedented shift towards online learning, compelling university students worldwide to engage in self-directed learning within remote environments. Despite the increasing importance of online education, the factors driving students’ motivation for self-directed online learning, particularly those involving economic incentives, have not been thoroughly explored. This study aims to address this gap by analyzing large-scale data collected from 19,023 university students across China during the pandemic. Using mixed Bayesian networks and multigroup structural equation modeling, the study explores the complex relationships between personal characteristics, academic characteristics, the academic environment, and students’ motivation for self-directed online learning. The results reveal significant associations between online self-directed learning motivation and personal characteristics. such as gender and age, academic characteristics, such as education level and learning incentives, and the geographic location of the school within the academic environment. Moreover, the causal relationship between school location and online self-directed learning motivation varies by gender and educational level. This research not only provides new empirical support for the theoretical framework of online learning motivation but also contributes to the broader fields of educational psychology and online learning research.

1. Introduction

In early 2020, the outbreak of the COVID-19 pandemic led to the largest educational disruption in history [1], and several countries and regions worldwide took measures to suspend classes in response to the outbreak. The effects of the outbreak did not spare extracurricular educational institutions in China; as universities went into lockdown, students could not participate in offline extracurricular training. Thus, after-school providers had to shift from traditional teaching models to online ones to reduce direct person-to-person contact [2,3,4]. In this context, the transition to online learning in higher education has accelerated, prompting a large number of university students to engage in self-directed learning within remote environments [5]. However, compared to primary and secondary school students, university students face the dual challenges of rapidly updating knowledge and increased employment pressures. To be able to keep up with the times and enhance their competitiveness, they usually take the initiative to study to acquire increasingly advanced knowledge and skills.

Proactive learning is not merely about meeting employment challenges, but it is also about achieving personal psychological fulfilment and self-worth. However, maintaining university students’ learning motivation, particularly in the context of online courses during the pandemic, has become a major challenge. Research has shown that different types of social support, professional courses, learning institutions, learning environments, and other factors affect self-directed learning motivation [6,7,8,9,10]. For example, Baeten et al. [6] studied the effects of different learning environments on students’ motivation to learn and achieve. Li et al. [2] found that students’ grade level, gender, and learning environment may be important factors affecting academic performance in online education, and Wei et al. [11] found that groups with self-directed learning motivation showed higher perceived learning outcomes than their peers.

While these studies have significantly contributed to understanding learning motivation in traditional educational settings, the factors driving university students’ self-directed online learning, especially those involving economic incentives, have not been sufficiently explored. Economic incentives, such as learning-incentive money (which may include rewards, subsidies, or scholarships), are particularly relevant in online learning environments. These financial incentives can motivate students by providing tangible rewards for academic achievements, which, in turn, can enhance their self-directed learning motivation. For example, monetary incentives have been shown to increase students’ engagement and effort in laboratory-based tasks, but their effectiveness in natural educational settings has yielded mixed results [12,13]. Understanding how learning-incentive money interacts with other personal and academic factors to influence motivation is crucial for developing effective strategies in online education.

Most research to date has focused on small sample sizes and has primarily analyzed offline learning models, making it difficult to generalize findings to the broader context of the global shift to online learning. Therefore, it is necessary to analyze data from large samples of university students during the COVID-19 pandemic to explore how personal characteristics, academic characteristics, and the academic environment influence online self-directed learning motivation. Additionally, conducting subgroup analyses among undergraduate, master’s, and doctoral students can provide deeper insights into how these variables affect different groups.

Based on this, the objectives of this study are: (1) to analyze how university students’ personal characteristics, academic characteristics, and academic environment variables influence online self-directed learning motivation using large-scale sample data, and to explore the causal relationships between these factors; (2) to conduct subgroup analyses of undergraduates, master’s students, and doctoral candidates to gain a deeper understanding of how these variables influence online self-directed learning motivation across different groups. The findings of this study will not only provide new empirical support for the theoretical framework of online learning motivation but also contribute to the fields of educational psychology and online learning research. Furthermore, these insights will offer valuable information and recommendations for the development of future online education platforms, course design, and teaching strategies, particularly in promoting the sustainable development of online education in a globalized context.

2. Theoretical Framework

2.1. Self-Determination Theory and Online Learning Motivation

Self-Determination Theory (SDT), developed by Ryan and Deci [14], is a widely recognized framework for understanding motivation within educational settings. SDT categorizes motivation into two types: intrinsic and extrinsic. Intrinsic motivation arises from the inherent enjoyment and interest in the activity itself, while extrinsic motivation is driven by external rewards or pressures. According to SDT, three basic psychological needs—autonomy, competence, and relatedness—must be satisfied to foster intrinsic motivation and optimal learning performance.

In the context of online autonomous learning, these needs are particularly critical. Autonomy refers to a student’s sense of control over their learning process, while competence refers to their confidence and ability to successfully complete online learning tasks. Research indicates that students with intrinsic or self-determined motivation typically achieve more positive learning outcomes [15,16]; for example, they are better able to adapt to academic life, have lower levels of perceived stress, and can engage in more sustained learning [17,18]. With the rapid advancement of technology, the manner in which people learn or acquire new knowledge has changed, and one major change is that an increasing number of students are choosing to receive education online [19]. Autonomy is a key feature in online learning [20,21]. This surge in demand for online learning resources further underscores the importance of self-directed learning motivation in this new educational environment.

2.2. The Impact of Personal Characteristics on Online Learning Motivation

Students’ personal characteristics, such as gender, age, and educational level, are closely related to their learning motivation [22]. A study found significant gender differences in self-directed learning motivation [20], where girls usually show a higher autonomous motivation to learn [23]; this trend may contribute to their higher academic achievement. Ethnicity is associated with student autonomy and, during distance learning throughout the pandemic, curiosity among Latino students enrolled in ethnic studies courses improved during the school year, although students’ stress and motivation were lower. Additionally, research has indicated that, as people age, they tend to develop a stronger sense of autonomy; older individuals are better at internalizing their goals and personal initiatives than younger individuals.

2.3. The Role of Academic Characteristics in the Formation of Learning Motivation

Academic characteristics, such as school atmosphere and reward systems [24], play distinct roles in influencing students’ motivation and online learning behaviors. Economic Incentive Theory posits that monetary rewards can significantly influence behavior by increasing the perceived benefits of specific actions [25,26]. In educational settings, economic incentives have been shown to enhance student engagement, particularly in environments where motivation may be low. In online learning, where immediate guidance and timely feedback are often challenging to provide, learning-incentive funds can serve as an effective external motivator, enhancing students’ extrinsic motivation and thereby promoting greater engagement in their studies [13]. It is worth noting that reward mechanisms in the form of external incentives may also lead to learner overdependence, with various implications for human behavior [27]. In addition, a correlation exists between academic climate and academic performance. Negative school belonging has been noted to negatively impact intrinsic motivation and perceived learning [28]. Higher-level schools offer a more comprehensive curriculum, better-quality educational resources, and a learning environment suitable for the development of self-directed learning skills. Such schools tend to encourage active inquiry and self-directed learning processes, and they may provide more external resources and opportunities to promote the development of self-learning skills.

2.4. The Impact of Academic Environment on Learning Motivation

The academic environment, including school type, geographic location, and economic background, often influences the resources and opportunities available for self-directed learning. Schools’ geographic location and economic level typically correlate with access to richer online learning resources. In areas of higher economic development, schools are likely to have more resources available for pedagogical innovation, thus providing students with a wealth of learning material and environments that enhance their self-directed learning. Conversely, in less-affluent economies, the investment capabilities of schools and families in education may be constrained, potentially inhibiting the promotion of self-directed learning. Students from lower socioeconomic backgrounds and rural communities may require additional support to overcome educational challenges—particularly those exacerbated by the COVID-19 pandemic. Furthermore, the nature of a school (i.e., whether it is academic, technical–vocational, or comprehensive) influences the development of self-learning capabilities. Academic-oriented schools may place greater emphasis on theoretical knowledge and research skill development, thereby facilitating self-directed learning, whereas technical vocational schools may prioritize skills-based and practical learning experiences. Overall, the evolution of self-learning is affected by multiple factors: a school’s type, nature, and geographic location can directly or indirectly impact the distribution of educational resources, selection of educational strategies, and improvement of educational quality, all of which, in turn, shape students’ capacity for self-directed learning.

This study integrates principles from Self-Determination Theory and Economic Incentive Theory to develop a comprehensive model that explains the factors influencing college students’ motivation for self-directed online learning. The model posits the following:

(1)

Learning-incentive funds, as an external motivating factor, can enhance extrinsic motivation while potentially fostering self-directed learning motivation in an online setting.

(2)

Personal characteristics (such as gender and age), academic characteristics (such as educational background and reward systems), and the academic environment (such as school type and geographic location) significantly impact students’ learning motivation, with these factors being particularly crucial in online learning environments.

(3)

The interaction among these variables is expected to reveal how different groups respond to the challenges and rewards of online learning.

3. Method

3.1. Data Collection

In our study, we used the online search frequency of students on the learning platform as a key indicator of online self-directed learning motivation. Online search frequency refers to the rate at which students actively search for learning resources and information through the platform during their studies. This behavior reflects students’ willingness and ability to actively acquire knowledge, making it an external manifestation of self-directed learning motivation [16,29]. Furthermore, online search frequency data is easily accessible and can provide real-time insights into students’ learning motivation, making it a feasible and meaningful indicator for large-scale studies.

The survey data collected during the pandemic was provided by the Chinese educational technology company Zhuolang Technology. This company offers a wide range of content across various academic disciplines, enabling both undergraduate and graduate students to easily access academic materials and learning resources. Additionally, the company provides learning incentives based on students’ academic performance, including rewards, subsidies, or scholarships. These financial incentives can be used by students to purchase learning materials, pay for course fees, or serve as rewards for completing specific learning tasks.

The raw data were collected in the form of web search logs, which detailed students’ search activities on the platform. Each log entry included information on search frequency, timestamps, and the types of content accessed. Additionally, demographic and academic information about the students, such as gender, age, ethnicity, educational level, and school characteristics, was included. To ensure the quality and accuracy of the data, we employed several data cleaning procedures. First, we removed any duplicate or incomplete entries from the dataset. Next, we standardized the format of the remaining data to ensure consistency across different variables. For categorical variables, such as search frequency, we categorized the data into five ordered levels based on the actual number of searches. Finally, we identified and appropriately handled any outliers that could impact the analysis, maintaining the integrity of the dataset. As a result, these data are highly reliable and representative, providing a solid foundation for analyzing and revealing differences and similarities between groups.

We utilized hybrid Bayesian networks (BN) and multigroup structural equation modeling (SEM) for the causal analysis of the data (Figure 1). In the first stage, we considered gender, age, and ethnicity as students’ personal characteristics, and school level, education level, length of time spent in school, and learning-incentive money as their academic characteristics. The number of learning incentives on the online education platform is the number of incentives students received on that platform. We used school type, school nature, school economic geographic location, and school regional geographic location as academic environment variables. We adopted a hybrid BN-based approach to identify causal links between online self-directed learning motivation and learner characteristics (student personal characteristics, academic characteristics, and academic environment variables) among university students. In the second stage, we extracted online self-directed learning motivation as the core key variable and built a multigroup SEM with high causal arc strength as a premise from which important causal relationships were mined. In the third stage, we used multigroup SEM for group difference analysis to determine whether significant differences existed between the groups.

3.2. Bayesian Networks

The BN is a probabilistic graphical model consisting of a set of random variables and a Directed Acyclic Graph (DAG) [30], where nodes denote random variables and edges denote dependencies between these variables. This model allows the dependencies encoded in the DAG to be mapped to the probability space through conditional independence relations, thus performing quantitative probabilistic inference under uncertainty. BN can effectively handle large sample data to determine causal relationships among variables related to online self-directed learning motivation. BN is capable of dynamically adjusting its network based on the data presented or input, enabling it to successfully manage large and multivariate datasets, as well as handle incomplete or uncertain knowledge or information [31,32]. However, BN is less robust in theoretical interpretation compared to SEM [33,34]. Therefore, some studies combine these two models to establish causal relationships [35,36].

3.2.1. Dependency Establishment

Let $\mathrm{U}=\left({\mathrm{x}}_1,{\mathrm{x}}_2,\dots ,{\mathrm{x}}_{12}\right)$ be the set of input points and a total of one variable which represent SEX, MON, FRE, LEN, EDU_L, AGE, SCH_L, NAT, ECO_L, LOC, TYP, SRN shown in Table 1. A BN on a set of variables U is a network structure which is a DAG on U and a set of probability tables [37]:

$$B_P=\left\{P\left(x_i|pa\left(x_i\right),x_i\in U\right)\right\}$$

(1)

where $pa\left(x_i\right)$ is the set of the antecedents of $x_i$ in BN and $i=1, 2, 3,\dots , 12$.

A BN represents joint probability distributions (2):

$$P\left(U\right)={{\displaystyle \prod }}_{x_i\in U}p\left(x_i|pa\left(x_i\right)\right)$$

(2)

3.2.2. Structure Learning for Weighted Hybrid Peter-Clark–Hill-Climbing (PC-HC) Algorithm

The Peter-Clark–Hill-Climbing (PC–HC) algorithm, based on a weighted mixture, takes full advantage of the independence of scores and conditions. Among them, the method based on the conditional independence test is more efficient and can obtain the optimal global solution [38]. The scoring algorithm is a local search strategy with the problem of getting caught in local poles and saddle points, but it is more flexible in the way it changes the network’s structure [39]. The hybrid algorithm first obtains the graph structure according to the Peter-Clark (PC) algorithm and then uses the Hill-Climbing (HC) algorithm to connect the learned structures of each subgraph, and weighted corrections are made to the two graph structures to obtain the optimal BN structure.

①

Peter-Clark algorithm (PC)

As an algorithm based on the assumption of conditional independence, the PC algorithm can be flexibly combined with various conditional independence tests to infer causality between variables and thus obtain the optimal global solution [40]. The core idea of the PC algorithm is conditional independence using the following d-separation principle: the set of nodes I can separate nodes m and n, and no valid path exists between m and n if and only if I is given.

②

Hill-Climbing algorithm (HC)

As a score-based structure learning algorithm, the HC algorithm usually starts with an empty graph. It explores the search space of the graph by adding, deleting edges, and edges to maximize the target score and stops iterating when changing the structure of the edge neighbors; then, the score no longer increases to obtain the optimal structure of the graph, which may also be a local optimum [41,42].

After obtaining network structures G1 and G2 based on the two methods, the new network structure G is obtained by mixing and weighting the networks using the set weight vectors and selecting the substructures for subsequent analysis.

3.3. Multigroup Structural Equation Model (SEM)

SEM is a causal modeling approach that combines causal information with statistical data to quantitatively assess the relationship between the variables under study [43]. It allows for better control of measurement errors than general regression analysis and supports the construction of complex multivariate models. The main purpose of multigroup SEM analysis test is to assess whether the path model map can be adapted for other different groups when matched to one group [44], as defined in [45].

$$\eta =B\eta +\Gamma \xi +\zeta$$

(3)

where $\mathsf{\eta }=\mathrm{m}\times 1$ is a vector of latent variables, in this project, $\mathrm{m}=6$ is reflected in the key network which supports the implementation of the structural equation model; $\mathsf{\xi }=\mathrm{n}\times 1$ is a vector of latent exogenous variables; $\mathrm{B}=\mathrm{m}\times \mathrm{m}$ is a matrix of coefficients associated with latent endogenous variables; $\Gamma =\mathrm{m}\times \mathrm{n}$ is a matrix of coefficients associated with latent exogenous variables; and $\mathsf{\zeta }=\mathrm{m}\times 1$ is a vector of error terms associated with endogenous variables.

Table 1. Descriptive analysis (n = 19,023).

Variables	Abbreviation	Count	Percentage (%)
Gender	sex
Men		12,333	0.65
Women		6690	0.35
Ethnicity	ETH
Ethnic Han		17,368	0.91
Ethnic Minorities		1655	0.09
Age	AGE
≤18		248	0.01
19		2686	0.14
20		4843	0.25
21		4464	0.23
22		3143	0.17
≥23		3693	0.19
School level	Sch_l
First-rate universities and disciplines		2979	0.16
Full-time college		11,535	0.60
Higher vocational college		4509	0.24
School-running nature	SRN
Public school		15,474	0.81
Private school		2630	0.14
Independent college		905	0.05
Chinese–foreign cooperatively run schools		14	0.00
Length of schooling (year)	LEN
1.5		3	0.00
2		838	0.04
2.5		68	0.00
3		5555	0.29
3.5		1	0.00
4		11,806	0.62
5		729	0.05
6		2	0.00
8		21	0.00
Type of school	typ
Research university		1432	0.08
Applied university		13,061	0.68
Technical university		4530	0.24
Education level	Edu_l
Junior college		5092	0.27
Undergraduate		13,050	0.69
Master’s degree		797	0.04
doctor		84	0.00
Online search frequency	fre
Low frequency		4400	0.23
Less frequency		3145	0.17
Average frequency		3859	0.20
High frequency		3806	0.20
Very high frequency		3814	0.20
School area geographic location	loc
South China		2352	0.12
North China		2886	0.15
East China		4816	0.25
Southwest		2238	0.12
Northwest		1451	0.08
Central China		3434	0.18
Northeast China		1846	0.10
Economic and geographic location of the school	Eco_l
First-tier city		7809	0.41
Second-tier city		4978	0.26
Third-tier city		3302	0.17
Fourth-tier city		2006	0.11
Fifth-tier city		928	0.05
Money (CNY)	mon
≤5		111	0.01
5–10		11,172	0.58
10–15		206	0.01
15–20		7191	0.38
≥20		343	0.02

Table 1. Descriptive analysis (n = 19,023).

Variables	Abbreviation	Count	Percentage (%)
Gender	sex
Men		12,333	0.65
Women		6690	0.35
Ethnicity	ETH
Ethnic Han		17,368	0.91
Ethnic Minorities		1655	0.09
Age	AGE
≤18		248	0.01
19		2686	0.14
20		4843	0.25
21		4464	0.23
22		3143	0.17
≥23		3693	0.19
School level	Sch_l
First-rate universities and disciplines		2979	0.16
Full-time college		11,535	0.60
Higher vocational college		4509	0.24
School-running nature	SRN
Public school		15,474	0.81
Private school		2630	0.14
Independent college		905	0.05
Chinese–foreign cooperatively run schools		14	0.00
Length of schooling (year)	LEN
1.5		3	0.00
2		838	0.04
2.5		68	0.00
3		5555	0.29
3.5		1	0.00
4		11,806	0.62
5		729	0.05
6		2	0.00
8		21	0.00
Type of school	typ
Research university		1432	0.08
Applied university		13,061	0.68
Technical university		4530	0.24
Education level	Edu_l
Junior college		5092	0.27
Undergraduate		13,050	0.69
Master’s degree		797	0.04
doctor		84	0.00
Online search frequency	fre
Low frequency		4400	0.23
Less frequency		3145	0.17
Average frequency		3859	0.20
High frequency		3806	0.20
Very high frequency		3814	0.20
School area geographic location	loc
South China		2352	0.12
North China		2886	0.15
East China		4816	0.25
Southwest		2238	0.12
Northwest		1451	0.08
Central China		3434	0.18
Northeast China		1846	0.10
Economic and geographic location of the school	Eco_l
First-tier city		7809	0.41
Second-tier city		4978	0.26
Third-tier city		3302	0.17
Fourth-tier city		2006	0.11
Fifth-tier city		928	0.05
Money (CNY)	mon
≤5		111	0.01
5–10		11,172	0.58
10–15		206	0.01
15–20		7191	0.38
≥20		343	0.02

4. Results

This section presents the main results of the study, with a detailed focus on how personal characteristics, academic factors, and academic environment variables influence college students’ motivation for online self-study. By analyzing large-scale sample data, we identify key relationships between these variables and online learning motivation, and examine the differences across various groups (e.g., gender, education level). These findings will serve as the foundation for the subsequent discussion, where we will further explore how these results address the research questions posed in this paper.

4.1. Descriptive Analysis

A total of 19,023 items of data for 2392 Chinese universities were obtained from the survey. According to the results of the descriptive statistical analysis (Table 1), men accounted for 65% of the sample size, and women accounted for 35%—slightly less than men. In terms of age distribution, students in their 20s predominated, accounting for 25% of the sample. In terms of school level, students from top-ranking and world-class universities accounted for 16% of the sample size; students from general universities accounted for 60%, and students from higher education institutions for 24%. In terms of the nature of schooling, most students were from public schools, accounting for 81% of the sample. In terms of student duration, most students (42% of the sample) were in four-year programs. In terms of the type of school, applied universities were the majority, accounting for 68% of the sample, followed by technical universities at 24% and research universities at 8%. In terms of students’ education level, undergraduate degrees were predominant (69% of the sample). In terms of the frequency of searches, this was very high for 20% of the sample and very infrequent for 23%. In terms of the schools’ regional geography, students were more evenly distributed, with 25% of the sample located in East China. In terms of the schools’ economic geography, most students were in first-tier cities, accounting for 41% of the sample.

4.2. Results of BN

In this study, a hybrid BN was adopted for structural learning to obtain causal links between variables, and the data were preprocessed using R 4.2.0 to obtain DAGs and key networks, as shown in Figure 2.

The results obtained using the PC and HC algorithms were generally consistent. However, to enhance the interpretability of the structure and make the results easier to understand and accept, we adjusted the network structure according to a priori knowledge; that is, we used a weighted integration method to reconstruct a more realistic hybrid BN (Figure 2b). Figure 2a is found to be more credible than Figure 2b based on domain knowledge, setting the percentage of network credibility of Figure 2a to 0.7.

Figure 3 shows the optimal BN for this study, where arcs with a scale close to 1 are considered more robust in terms of causality, whereas the opposite is true for arcs close to 0. We extracted substructures using online self-learning dynamics as key variables to provide a basis for subsequent group difference analysis and obtained the following key network using dual core search frequency and learning-incentive money as leaf or parent nodes (Figure 3). The independent learning variables are mainly causally related to gender and age in the students’ personal characteristics, education level in the students’ academic characteristics, and the school area’s geographic location in the academic environment variables. In contrast, in this study, the independent learning variables were not related to ethnicity, school level, year, school type, nature of school operation, and economic geographic location of the school.

4.3. Results of SEM

After the BN combined a causal model that gave the maximum likelihood based on the data, we used SEM to check which nodes and arcs were reasonable and determine the robustness of the BN’s bootstrap approach. In addition, to enhance the interpretability of the structure and make the results more understandable and acceptable, we slightly adapted the initial structure to the domain knowledge.

Table 2. The fitting index value of multigroup structural equation modeling (n = 19,023).

	X2	X2/df	RMSEA	IFI	CFI	GFI
No-group measured value	66.151	9.450	0.021	0.951	0.950	0.999
Gender
Women	5.34	0.995	0.025	0.972	0.972	0.998
Men	14.999	3	0.013	0.978	0.978	1
Education level
Junior college	9.253	0.313	0.016	0.972	0.971	0.999
Undergraduate	19.799	4.95	0.017	0.971	0.971	0.999
Master’s degree	9.968	2.492	0.041	0.524	0.088	0.995

presents the final model fit indices. The root mean squared error of approximation (RMSEA) [46] is less than or equal to 0.08; the incremental fit index (IFI) is greater than or equal to 0.90; the comparative fit index (CFI) [47] is greater than or equal to 0.90, and the Goodness of Fit Index (GFI) is greater than or equal to 0.90, which is considered to be a good model fit [48]. The important fit indicators in this study are within the acceptable range of suggested values; therefore, the SEM model fits well.

Figure 4 shows the standardized path coefficients and significance levels between the variables in the study model. The SEM results largely confirmed the significant arc of BN, except that student search frequency was negatively correlated with school area and geographic location. In addition, the path coefficients for the total sample and the subgroups converged broadly, demonstrating stable causal pointing between the variables. Of these, in terms of motivation to learn independently online, online search frequency (β = 0.082, p = 0.001) significantly influenced learning-incentive money. Gender (β = 0.038, p = 0.001) and age (β = 0.135, p = 0.001) also significantly influenced the frequency of online searches.

4.4. Results of Multi-Group SEM

To determine whether the associations between subjective learning motivation and college students’ age and geographic location of the school area differed across gender and education level, we used two variables—college students’ gender and education level—as the basis for multigroup analysis to divide the model. The results (Table 2) showed that the important fit indicators were within the acceptable recommended values; therefore, the model fit was good.

Figure 5a shows the results of the multigroup SEM analysis based on gender. The SEM for women (Figure 5a) shows a positive and significant correlation between age and students’ search frequency (β = 0.128, p = 0.001) and education level (β = 0.291, p = 0.001), a positive and significant correlation between students’ search frequency and learning-incentive money (β = 0.085, p = 0.001), and no correlation between students’ search frequency and school area’s geographic location (β = −0.016, p = 0.178); moreover, education level (β = 0.042, p = 0.001) was positively and significantly associated with learning-incentive money.

The SEM for men (Figure 5b) showed a positive and significant correlation between age and students’ search frequency (β = 0.139, p = 0.001) and education level (β = 0.087, p = 0.001), a positive and significant correlation between students’ search frequency and learning-incentive money (β = 0.078, p = 0.001), and a negative and significant correlation between students’ search frequency and school area geographic location (β = −0.024, p = 0.008); moreover, education level (β = 0.053, p = 0.001) was positively and significantly associated with learning-incentive money.

Then, we analyzed multi-cluster structural equation modeling based on education level. Figure 6 (right) shows the results of the multigroup SEM analysis based on education level. The model plots for students in junior college education and the individual path coefficients (Figure 6a) show a negative correlation between students’ search frequency and school area’s geographic location (β = −0.033, p = 0.020), a negative correlation between genders (β = 0.036, p = 0.010), a positive correlation between age (β = 0.127, p = 0.001) and students’ search frequency, and a positive and significant correlation between students’ search frequency and learning-incentive money (β = 0.080, p = 0.001).

The model plots for undergraduate students and the individual path coefficients (Figure 6b) show that students’ search frequency is not related to the school area’s geographic location (β = −0.015, p = 0.083); gender (β = 0.041, p = 0.001) and age (β = 0.168, p = 0.001) are positively related to students’ search frequency; and students’ search frequency is positively and significantly related to learning-incentive money (β = 0.086, p = 0.001).

The model plots for students with postgraduate qualifications and the individual path coefficients (Figure 6c) show that students’ search frequency was independent of school area geographic location (β = −0.053, p = 0.117); gender (β = 0.024, p = 0.476) and age (β = 0.055, p = 0.101) were independent of students’ search frequency; and students’ search frequency was independent of learning-incentive money (β = −0.005, p = 0.874).

The findings for the postgraduate group are contrary to those of the overall relationship graph. Potential influences may cause the relationships for the variables explored to not hold, and this group may be more oriented toward academic studies than toward traditional exam education. Therefore, the learning characteristics explored did not have a significant causal relationship in this group.

5. Discussion

In terms of students’ personal characteristics, gender and age were significant factors influencing motivation for autonomous learning. Furthermore, age had a slightly higher impact on self-directed learning in the men’s group than in the women’s group, which may be due to the traditional division of labor between men and women, where men are primarily responsible for providing for their families. Although the number of women has risen in recent years, and their contribution to household income has grown, overall, the core of household income remains male. Under the same social pressures, older men, who are often pressured to be employed to support their families, may be more anxious about learning independently (and more motivated to do so) than women. In addition, the effect of age on educational attainment is significantly more relevant in the men’s cohort than in the women’s cohort, which may be due to the fact that age plays a crucial role in women’s post-college human capital investment decisions. Zhang et al. note that, according to Census and CFPS data, women who enter school at a relatively later age are less likely than men to enter graduate school [49]. The act of delaying employment and marriage as they age puts women at a disadvantage in the labor and marriage markets, which is considered an opportunity cost of investing in human capital after college.

In terms of students’ academic characteristics, learning-incentive money was positively correlated with motivation for autonomous learning. This is consistent with previous research showing that small rewards enhance autonomous motivation [50]. Learning-incentive money is a moral and material incentive to help students realize their self-worth, which motivates college students to study independently online and prompts them to engage more actively in their own behavior. Appropriate rewards can replace feedback from online educators, making it possible to achieve desired learning outcomes in a learning environment with limited personal interaction. The education level was also significantly and positively associated with learning-incentive money; this may be because less-educated students have less awareness of their careers, lack initiative and self-motivation to learn, and therefore receive less learning-incentive money. In addition, students with specialized qualifications tend to have a pronounced herd mentality and generally less self-control compared to undergraduate students and are prone to indulging in these establishments, which are conveniently located and rich in recreational venues [51]. Therefore, a region’s geographic location influences the self-learning motivation of students with tertiary qualifications. The doctoral degree group is more inclined to academic research than traditional examination-based education; therefore, in this group, we found no significant causal relationship among the learning characteristics explored.

In terms of academic environment, students’ motivation for self-learning was often affected by the geographic location of the school area. For example, the level of Internet development in China’s eastern coastal regions was significantly higher than that in the central and western regions, and the level of network construction in cities was significantly higher than that in rural areas. The existence of poorer network environments in remote areas, the lag phenomenon when learning online, and the inability to operate and experience technological software all directly affected the students’ learning progress and search results. We also found that the women’s and men’s groups diverged in terms of the impact of the school area’s geographic location on self-directed learning, and this gender difference may be related to cultural and social expectations. According to Ong, cultural, social, and institutional biases contribute to women’s inability to fit into the current culture’s power spaces, which are more pronounced in science, technology, engineering, and mathematics (STEM) professions [52]. Some cultures tend to place more emphasis on men’s roles in the workplace and society, whereas women are more often considered the primary caregivers of the family. This cultural and societal expectation may lead men to focus more on advantages of geographic location, such as school reputation and ranking.

6. Conclusions

University students’ education is essential for the dissemination of knowledge, technological innovation, and sustainable social development. Although the COVID-19 pandemic has transformed traditional offline classroom education into online education, higher education still faces new challenges and opportunities. Taking Chinese college students as the research object, we selected factor variables related to their online learning, and models and analyzed their motivation for online self-directed learning by means of hybrid BN and multigroup SEM. The main findings and conclusions are summarized as follows.

Overall, the SEM test results were broadly consistent with those of the mixed BN key network and multiple SEM subgroups, suggesting a stable causal relationship between the factors. Self-learning motivation was mainly related to gender and age in students’ personal characteristics, to education level in students’ academic characteristics, and to the school area’s geographic location in the academic environment variables. Learning-incentive money had a significant positive effect on self-directed learning motivation for different gender groups; more specifically, it was not related to the school area’s geographic location in the women’s group, whereas it had a negative correlation with it in the men’s group. Among groups with different education levels, the self-learning motivation of students with specialized degrees had a negative correlation with the school area’s geographic location, whereas the online search frequency of students with bachelor’s degrees was not related to the geographic location of their school area. These findings help to improve the understanding of college students’ motivation for online self-directed learning and recognize the future role of online educational environments in promoting health. Based on the study’s findings, future online education platforms can offer personalized learning incentives tailored to students’ gender, age, and educational background, thereby enhancing their self-directed learning motivation. This study not only expands the current understanding of how demographic characteristics, economic incentives, and educational factors influence learning motivation, providing a foundation for further research in educational psychology and online learning, but also contributes to the sustainable development of online education in a globalized context.

7. Limitations and Future Prospects

The study identified several key factors that influence students’ self-directed learning motivation in online environments, including gender, age, educational background, and the impact of various incentives. These findings offer valuable insights for the design and development of future online education platforms.

7.1. Personalized Learning Incentives

Based on our research results, future online education platforms can implement personalized learning incentives tailored to the specific characteristics of students. For example, gender-specific motivational strategies can be developed to address the unique learning preferences and challenges faced by male and female students. This is particularly important given the study’s findings that gender plays a significant role in self-directed learning motivation. Similarly, age-appropriate rewards can be designed to align with the developmental stages and cognitive needs of different age groups. Additionally, rewards suited to students’ educational backgrounds can be created to support varying levels of academic preparedness and experience, thereby enhancing self-directed learning motivation. By adopting these targeted approaches, online education platforms can more effectively engage students and foster a more productive learning environment.

7.2. Tiered Learning Experiences

The study also highlights the different motivational needs across educational levels (undergraduate, master’s, doctoral). Considering the volume of learning and research tasks at each stage, course designers and instructors are advised to implement tiered online learning experiences for undergraduates. For example, a virtual research community could be created for master’s students, including additional learning sections on research submissions and writing, serving as a reliable source for academic support, enabling students to receive timely feedback from their peers.

7.3. Comprehensive Support Systems

Recognizing the importance of various academic characteristics and environmental factors identified in the study, system developers can create multiple channels to facilitate students’ help-seeking behavior, such as online Q and A forums or collaborative help-seeking tools embedded in learning management systems [49,51]. Platforms can offer different types of rewards, such as incentives for users to reopen the app, significantly increasing the likelihood of repeated use. These platforms could combine rewards, like points, levels, badges, tasks, and leaderboards, with visual feedback to encourage ongoing participation and enhance students’ motivation for online self-directed learning [52].

7.4. Limitations

The ability to learn independently online is influenced by many other factors, such as different online search models, teachers’ teaching methods, the online platform’s management style, students’ personality and behavior, and their family’s educational background. Furthermore, as the participants in this study were drawn from various disciplines, the results derived can only provide general insights into the online learning profiles of university students. Students from specific fields or domains are likely to exhibit learning tendencies that differ from those identified in this study and necessitate further research. Therefore, in the future, we intend to extend this analysis to include additional data from other large online-learning courses across various disciplines. In addition, other influencing factors, such as learner time interactions, different online search patterns, teachers’ teaching-method preferences, and family educational background, will be utilized to present a relatively comprehensive set of models for detailed analysis.