1. Introduction
As digital transformation progresses in the workplace, employees engage intensively with mobile platform-mediated work, reshaping how work activities are executed. The omnipresence of smart devices, such as mobile phones and tablets, has driven the development of mobile versions of work-related systems, making work more flexible, autonomous, and porous. The widespread adoption of mobile platforms exemplifies successful digital transformation within organizations, as evidenced by their high popularity and extensive user coverage across various industries. By the end of 2022, DingTalk amassed 600 million users worldwide and over 23 million enterprise organization accounts [
1]. Similarly, Slack had an estimated 35 million active users in 2022 [
2].
In the digital era, the increasing integration of work and life has become apparent, with employees finding it challenging to maintain a clear boundary between the two. Consequently, mobile platforms face criticism for allowing work to encroach on personal lives. These concerns burden organizations striving for sustainable digital transformation and make decision-makers more apprehensive about implementing new technology.
As described in [
3], the mobile platforms in this research refer to digital platforms that support remote working, particularly by facilitating communication and collaboration. Examples include Slack, DingTalk, and Microsoft Teams. Mobile platforms enable individuals to work from anywhere at any time. However, despite their positive influences on productivity, the increased connectivity and flexibility afforded by these platforms are often criticized in academic scholarship for leading to adverse health effects [
4,
5], work–life imbalance [
6], and work exhaustion [
7]. These documented negative consequences raise public concerns about the sustainability of these technologies and whether the gains in productivity come at the cost of personal well-being [
8]. Moreover, these platforms render work more intrusive, causing individuals to be perpetually surrounded by work activities, ultimately fostering a perception that technology is detrimental. These narratives predominantly emphasize the negative impacts of mobile platforms, thereby supporting the assertion that technology is the primary culprit behind work–life imbalance (WLI).
Work–life imbalance (WLI) can be conceptualized as the phenomenon of working excessive hours, where work encroaches upon personal lives [
9]. Empirical studies indicate that the working population in Asia, including China, is typically subjected to a long-hour work culture. China in particular has garnered significant attention for the “996” working schedule—working from 9 a.m. to 9 p.m., six days a week—which has become a contentious issue among both employees and scholars.
This pervasive work culture significantly impacts work–life balance (see [
10,
11]). Coupled with the rapid pace of organizational digital transformation, it can be inferred that Chinese employees experience a technology-induced imbalance between their work and personal lives. Notably, this imbalance is often primarily attributed to the ubiquity of mobile platforms, which are seen as a main culprit in engendering employee work–life imbalance. This prompts us to consider the actual role that mobile platforms play in causing WLI in the Chinese workplace.
In many real-world situations, the relationships between variables are complex and often follow nonlinear patterns resulting from interactions between variables and threshold effects. Nevertheless, existing studies frequently examine the linkage between mobile platforms and work–life imbalance using linear regression methods. This raises doubts. Are the results sufficient to explain human behavior by several variables with linear assumptions? Ignoring nonlinearities in data can lead to inaccurate models and predictions, thereby missing opportunities to uncover critical insights. Human behavior is inherently more complex; however, to date, few studies have employed alternative methods such as machine learning algorithms to address this topic.
As workplaces increasingly integrate Artificial Intelligence (AI), it becomes imperative that techniques for user behavior detection and prediction employ advanced methodologies consistent with AI-driven practices. To adeptly navigate this transformative trend towards a sustainable future, it is essential to apply AI-related methods to comprehensively understand the role of technology in shaping individuals’ work–life integration behaviors. Given the complexity of technology-influenced human behavior and the relative paucity of AI techniques employed in predicting work–life imbalance, our objective is to employ the explainable AI framework to elucidate how various determinants affect work–life imbalance, particularly when employees engage with pervasive mobile platforms.
Our research questions are straightforward:
What are the complex nonlinear relationships between mobile platforms and work–life imbalance, and should technology be held entirely accountable for invading private lives? Are there other associated factors?
How does the application of the explainable AI framework enhance our understanding of the role of mobile platforms in shaping employees’ work behavior?
How can the obtained evidence provide actionable guidance for the sustainable success of mobile platforms?
This paper extends the research frontier and makes several significant contributions. First, it contributes to the literature by examining the role of mobile platforms in influencing individuals’ integration behaviors through the lens of explainable AI. This approach supports decision-making processes and promotes a sustainable future amidst the digital transformation. Second, building on the extensive discussions in existing studies on the causes of work–life imbalance, we identify the input features for our predictive model using the co-occurrence network. This methodology allows us to derive conclusions without relying on prior hypotheses, thus revealing intricate facts free from knowledge bias. Third, the proposed hybrid method, which integrates the co-occurrence network with the explainable AI framework, not only provides robust conclusions with deeper insights but also introduces a novel methodological scheme. This advances the intersection of management and AI applications.
The remainder of this paper is structured as follows: In
Section 2, we review the related background. Our methodology encompasses two sub-studies.
Section 3 presents the first sub-study, including feature extraction and corresponding results. In
Section 4, we discuss the second sub-study, which includes the explainable AI framework and its results.
Section 5 offers a discussion of the findings and contributions. Finally,
Section 6 provides conclusions and suggestions for further research.
2. Related Background
2.1. Work–Non-Work Integration and Boundary Theory
Boundary theory is central to a wide range of work–life integration (WLI) publications that explore its causes and consequences. According to boundary theory, individuals are boundary crossers who switch roles between work and family domains, incorporating both psychological states and physical environments [
12]. When individuals continue to engage in work-related activities within the family domain (e.g., thinking about work-related matters), they are unable to effectively perform their roles in the family domain (e.g., cultivating personal interests and hobbies) [
13,
14]. This process is referred to as boundary permeation.
Conceptually, boundary permeation occurs when an individual participates in a role within one domain under physical conditions while simultaneously engaging in a role in another domain under psychological conditions or through actual behavior [
15]. As this phenomenon has directionality, it can be represented as unidirectional (work permeating into life or life permeating into work) or bidirectional (work permeating into life and life permeating into work) [
16].
Undoubtedly, mobile platforms break down the boundaries between work and life domains, enabling individuals to integrate their work into their private lives in an extremely convenient manner. In this context, technology-induced integration behavior is defined as the behavior of integrating from work to non-work domains, rather than the reverse. The most significant function of mobile platforms is their strengthened connection with work. Following this logic, existing studies argue that increased use of mobile platforms correlates with increased platform-mediated work demands [
17,
18], extending regular work hours and subsequently leading to work–non-work integration behaviors.
2.2. Explainable AI Framework for Behavior Prediction
Artificial Intelligence encompasses a wide array of disciplines, with machine learning being one of the most prominent tools for addressing complex problems and enhancing predictive accuracy. Although conventional machine learning algorithms effectively solve complex issues, their results often lack interpretability due to the opacity of their processes. To address this challenge, Shapley Additive Explanations (SHAPs) have been introduced as a powerful tool to enhance the transparency of algorithmic outcomes. SHAPs provide “shapley values” to assess the importance of features contributing to integration behavior. The integration of machine learning algorithms with SHAP techniques establishes an explainable AI framework [
19,
20]. Following existing literature, we apply this framework to our context to uncover the true relationship between determinants and work–life imbalance. This explainable AI framework not only enables the identification of nonlinear effects behind technology-induced imbalances but also helps differentiate patterns of integration from non-integration behaviors.
Machine learning (ML) possesses advanced predictive capabilities, providing profound insights and facilitating decision-making across a wide range of areas. For instance, XGBoost has proven effective for earthquake prediction and warning [
21,
22]. In study [
23], the authors employed the explainable AI framework for real-time accident detection. Another study utilized ML to predict driver fatigue, aiding the transition from manual to automated driving [
24]. Additional applications include predicting choices related to electric vehicle charging stations [
25] and intercity travel preferences [
26]. These studies, embracing the advances in artificial intelligence and machine learning, offer nuanced insights and solutions for real-world problems. There is widespread consensus that ML excels in detecting complex non-linear relationships and uncovering hidden patterns in data, unlike traditional statistical models such as linear regression and generalized linear models, which rely on prior assumptions about the data.
2.3. Breaking away from Linear Regression Method
We present
Figure 1 to illustrate the complexity of human behavior in real-world scenarios, emphasizing the necessity of employing machine learning techniques to predict employee integration behavior. Given the intricate nature of human actions, the function that most accurately captures behavior dynamics is likely quadratic or trigonometric, as depicted in the central section of
Figure 1. The determinants of behavioral decisions, specifically in the context of work–life integration, encompass heterogeneous factors from various dimensions, including organizational and personal levels. Outliers may exist among numerous observations, and the relationships between independent and dependent variables are often nonlinear, manifesting in curved influence patterns, as shown in
Figure 1.
Evidence 1 is derived from linear assumptions among variables, whereas Evidence 2 stems from curvilinear assumptions. By comparison, it is evident that Evidence 2 more closely approximates real human behavior, as it models human behavior as a dynamic system with multiple peaks and troughs over different intervals. This suggests that, within interval A, the relationship may exhibit a stable upward trend, whereas in interval B, it may shift to a downward trend. A substantial body of existing research employs traditional linear assumptions toward employee behavior; however, recommendations based solely on this method are likely to be misrepresented, failing to reveal the true nature of employee behavior and providing inaccurate decisions to support digital transformation.
In summary, we propose that work–life integration behavior is likely to be intricate and dynamic, which cannot be adequately addressed by linear regression methods, as illustrated in
Figure 1. During an organization’s digital transformation, understanding the subtle impacts of smart platforms and work–life integration is critical. The existing paucity of curvilinear behavior exploration, combined with the limitations of linear regression methods, compels us to develop a novel approach to thoroughly revolutionize the understanding and prediction of employee behavior in the digital era.
As previously discussed, machine learning techniques hold significant promise for addressing complex behaviors and aiding decision-making processes. To this end, we propose a novel data-driven method comprising two sub-studies. In Study 1, our objective was to consider as many potential influencing factors as possible without theoretical biases. Specifically, by calculating the co-occurrence frequency in the existing literature, we identified potential determinants co-appearing with work–life integration behavior. Once the relevant features were confirmed, we distilled each feature into specific constructs with measurable items and subsequently converted them into input variables for behavioral prediction.
In Study 2, we developed a predictive model based on machine learning algorithms to forecast work–life integration behavior. We used survey data as the input for the predictive model and employed the SHAP (Shapley Additive Explanation) technique to elucidate the underlying mechanisms of work–life integration behaviors. In addition to the results obtained from machine learning and SHAPs, we meticulously conducted scale validation and statistical assessments before generating the machine learning models. To the best of our knowledge, the combination of survey data and machine learning has seldom been verified, underscoring the novelty and contribution of our study.
3. Study 1: Feature Extraction Based on Co-Occurrence Network
3.1. Methodology
We commence our investigation with a comprehensive literature search using the Web of Science database. Employing the keywords “work–life boundary”, “work–life conflicts”, and “work–life integration”, we identified 578 articles published between 2018 and 2023. Our review was meticulously focused on the determinants of integration behaviors rather than their consequences. Additionally, given our interest in technology-induced work–life imbalances, we excluded general topics that did not explicitly address mobile technologies or ICTs. Ultimately, 73 highly relevant articles were retained for feature extraction. To illustrate our procedure, we selected a pertinent example article (e.g., [
27]) that discusses technology-mediated interruptions and their impact on employees’ work–life boundary experiences for subsequent content analysis.
Following the selection of articles, Atlas.ti (version 8.4.3) was utilized to identify key phrases. This software is widely used in qualitative research for text analysis [
28] and was therefore adopted to extract features from the retrieved articles. While the standard procedure for employing Atlas.ti in qualitative analysis involves coding, conceptualization, and network analyses [
29], we primarily used the software to extract key phrases from the introduction and literature review (or theoretical background) sections. Given that the selected articles are well-organized empirical analyses that often explicitly state relevant concepts or factors, we extracted key phrases directly from the authors’ arguments. For instance, in a sentence from [
27], “Some individuals prefer to segment work and non-work domains through thicker boundaries, whereas others may prefer more porous boundaries that integrate work and non-work domains”, we identified and coded key phrases: “preferences”, “porous boundaries”, “thicker boundaries”, and “work–non-work domains”.
In the subsequent phase, the co-occurrence network technique was utilized to construct the network. A co-occurrence network is a critical tool in text mining and Natural Language Processing (NLP), capturing the frequency with which pairs of phrases appear together within our lexical dataset. Each entry in this network indicates the frequency of a particular pair of phrases co-occurring in the predefined context, providing a structured method to analyze textual data. The Python software (Python 3.8.19 in Spyder IDE 5.5.1) was used to visualize the co-occurence network. The pseudocode for drawing the co-occurrence network and the mathematical expression of calculating the co-occurrence matrix are provided in Algorithm A1 in
Appendix A and Equation (
A1) in
Appendix D, respectively. In essence, this method allows us to identify and quantify the frequency of phrase co-occurrence, facilitating the process of determining the key factors influencing work–non-work integration behavior.
3.2. Results
3.2.1. Potential Influencing Features
Figure 2 depicts the nodes representing frequencies and the top five clusters detected in the co-phrase network. The node size reflects the frequency of phrase appearances in the lexical dataset. Notably, constructs such as ICT-enabled availability and constant connectivity are frequently mentioned in discussions on work–life integration behavior, aligning with prior findings. Boundary tactics also emerge as potential influencers of integration behavior. In consultation with information science experts and supported by the literature, we manually restructured and color-coded these clusters. The identified clusters include Availability, Autonomy and Control, Boundary Preference, Norms, Role Centrality, and Integration Behavior. The explanations for each cluster are detailed in
Table 1.
Based on
Table 1, we identified key features influencing work–non-work integration and mapped them into measurable variables for data collection and model training. We selected 16 relevant articles that provide measurements of variables and examine their relationships with the target variable. This process enabled us to refine clusters from the literature into specific variables that serve as input features for machine learning models. The mapping from clusters to variables and their definitions are detailed in
Table 2.
3.2.2. Survey Administration and Data Collection
Upon completing the feature extraction process, we identified the variables necessary for constructing the predictive model and measures for data collection. We emphasize that the survey method employed in this study significantly differs from traditional hypothesis-driven linear regression approaches. First, we do not propose any hypotheses; instead, we identify factors that may influence the results. Second, we consider all these factors as input features for predicting the target variable and disregard mediation, moderation, or direct linkages commonly used in traditional survey-based research. We focus solely on whether a factor increases or decreases the behavior and in which direction.
Consequently, our method, although survey-based, is inherently a hybrid approach that bridges traditional managerial studies with machine learning techniques. During the survey stage, measurement model validation is necessary; hence, statistical examinations were conducted to ensure that the survey items accurately reflect the latent variables. For structural model validation, coefficients of the path estimates are discarded. Since the data are primarily used for ensemble learning techniques, it is redundant to test whether the constructs are theoretically discriminant or related, which are the main purposes of convergent and discriminant validity. For details on the advantages of ensemble learning, please refer to existing articles in Computer Science applications [
35,
36]. In the machine learning stage, the research design adheres to paradigms established in CS applications by presenting the algorithms used in the
Section 3.1. This is followed by model comparison, model selection, hyperparameter tuning, and SHAP interpretation [
37,
38]. Details regarding this are explained in Sub-study 2.
The entire data collection process was meticulously conducted following established behavioral science research paradigms. To ensure content validity, measures were adapted from existing literature. A rigorous translation/back-translation procedure was employed to ensure semantic equivalence in both Chinese and English. The input variables included Available Demands on Mobile Communication Platforms (AVD), Segmentation Preferences (SEPs), Boundary Control (BOC), Work Identity (WOI), and Integration Norms (INNs). The dependent variable is Work–Non-Work Integration (WWIN). These measures demonstrated robust psychometric properties in previous research. Detailed items for the measure are provided in
Table A1 in
Appendix B.
The preliminary investigation was carried out in May 2023, during which we recruited a small set of participants through a summer training campaign tailored for public employees. While we did not prioritize specific industries, we expected participants to be familiar with various mobile platforms to ensure they had direct experience with technology-related work–life (im)balance. The professional backgrounds of the participants included public officials with long-term appointments, employees of third-party social welfare organizations, and members of state-owned financial agencies, among others. The preliminary investigation indicates that these participants worked in front-line positions heavily utilizing digital tools, which is well aligned with our research objectives, thus affirming their representativeness.
We continued to use the convenient sampling method to collect a larger dataset. From May to August 2023, questionnaires were disseminated to a broader pool of participants, also from the training courses, ultimately reaching 840 employees. Academic rigor and ethical standards were strictly adhered to throughout the procedure. Socioeconomic factors were collected using parsimonious rules. First, anonymity was guaranteed, and the purpose of the survey was clearly communicated to all participants both verbally and through written explanations on the questionnaire. We emphasized that the collected personal information would only be used for statistical and research purposes, with no inclusion of sensitive private data. Second, participation was entirely voluntary, with full respect for each individual’s willingness. Third, participants were required to be familiar with mobile technologies, such as work apps; otherwise, they would not satisfy the research criteria.
Out of the 834 responses obtained, 176 were excluded due to incomplete answers, the same propensity for all answers, or inattention. Consequently, 658 valid responses were included in the subsequent analysis, resulting in a valid response rate of 79%.
4. Study 2: An Explainable AI Framework for Predicting Nonlinear Behavior Patterns
4.1. Methodology
4.1.1. Linear Regression, Random Forest, LightGBM, and XGBoost
To predict the integration behavior, we start with the simplest algorithm known as Linear Regression (LR). LR employs ordinary least squares to estimate the linear relationships between the variables. We let
be a feature mapping from the input space
to
and consider the family of linear hypotheses:
Linear regression consists of seeking a hypothesis in with the smallest empirical mean squared error.
Although LR is easy to implement and understand, it lacks a robust generalization guarantee. This is because it focuses solely on minimizing empirical error without regulating the weight vector’s norm or incorporating other forms of regularization. LR suffers from various statistical restrictions, such as endogeneity and multicollinearity; thus, its performance is typically poor in most applications.
Random Forest (RF) operates on an ensemble learning paradigm, first proposed in [
39]. In the RF model, a variety of decision trees are developed using different subsets of the data, and each decision tree is trained using the bootstrap strategy. For regression, the RF model averages the predictions of all
B decision trees. The final prediction
for an input
x is given by
For classification, the RF model aggregates the predictions of all
B decision trees by majority voting. The final prediction
for an input
x is given by
The RF excels in mitigating the potential overfitting issues, with high prediction accuracy and stability.
The LightGBM (Light gradient boosting machine) is a highly efficient and scalable implementation of gradient boosting decision tree (GBDT) algorithms, designed for fast training speed and high efficiency [
40]. GBDT is an ensemble learning method that combines the predictions of multiple weak learners, typically decision trees, to produce a strong learner. The objective of GBDT is to minimize a loss function
by iteratively adding decision trees to the model. The prediction at iteration
m is given by
where
is the prediction from the previous iteration,
is the learning rate, and
is the new decision tree added at iteration
m. Two main important features of LightGBM are gradient-based one-side sampling (GOSS) and exclusive feature bundling (EFB), which contribute to the computational efficiency with an acceptable level of accuracy. LightGBM is particularly well suited for large datasets and can handle high-dimensional data with ease. Therefore, processing a small dataset via LightGBM might cause overfitting. More detailed explanations of GOSS and EFB techniques and applications of LightGBM can be found in [
41,
42].
XGBoost (Extreme Gradient Boosting) is an advanced implementation of gradient boosting algorithms designed to be highly efficient, flexible, and portable. It provides parallel tree boosting that solves many data science problems in a fast and accurate way [
43]. The core idea behind XGBoost is to use a gradient boosting framework, which builds models sequentially, each new model correcting errors made by the previous ones. The effectiveness and efficiency of XGBoost have made it a popular choice in various machine learning competitions and practical applications [
44]. The objective function to be minimized in XGBoost is composed of a convex loss function
L that measures the difference between the prediction
and the target
, and a regularization term
that penalizes the complexity of the model:
where
represents the parameters of the model,
n is the number of data points,
K is the number of trees, and
represents the individual trees.
is a regularization term to reduce overfitting, which can be expressed as Equation (
6),
where
T represents the number of leaf nodes in the
kth decision tree,
and
can decide penalty strength together,
is the score of the
ith leaf node.
XGBoost is one of the most popular and powerful machine learning algorithms given its robustness and accuracy in both small and large datasets [
45,
46]. In additon, we emphasize that XGBoost is not impacted by multicollinearity [
47], which is a common concerning issue in utilizing surveys as data sources. Reflecting on our case, even if the two variables in the survey are in resemblance and capture similar perceptions, both can be kept and would not conflate the significance of different features. Prediction performance based on these algorithms is shown in the following sections.
All these algorithms are implemented through the Python programming software (Python 3.8.19 in Spyder IDE 5.5.1).
4.1.2. Hyparameter Tunning
Hyperparameter tuning is performed using K-Fold Cross-Validation (CV), a traditional search strategy employed in various stages of the machine learning process, including hyperparameter tuning, feature selection, and model building. In K-Fold CV, the training dataset is randomly divided into
K equal-sized complementary subsets. These subsets are used in rotation to train the model, with one subset held out for validation in each iteration. Common choices for the value of
K are 5 or 10, although no formal rule dictates this selection. The “GridSearchCV” function in the “sklearn” package provides an effective grid search algorithm for identifying optimal hyperparameters. In this study, the training dataset is partitioned into ten folds (
), as recommended by previous research [
48]. Random subsampling is employed to isolate the data from the entire sample by dividing it into
K subsets. For each data split, a fixed number of observations from the sample are retained as experimental data without replacement. Subsequently, for each split
K, the predictive model is trained from scratch on the training data, and the prediction error for each test set is calculated. The example below demonstrates the implementation of K-Fold CV using Python (Python 3.8.19 in Spyder IDE 5.5.1). For brevity, the hyperparameter settings and model training steps are omitted.
kf = KFold(n_splits=10, shuffle=True, random_state=42)
grid_search = GridSearchCV(
estimator=model,
param_grid=hyperparameters,
cv=kf,
scoring=’accuracy’,
verbose=2
)
4.1.3. Shapley Additive Explanations
SHAP is a unified framework for interpreting machine learning models. The combined effect of multiple independent and dependent variables can be explained and quantified by employing the Shapley value. The core concept behind the Shapley value is to calculate the mean value of the marginal contribution of a feature to the model prediction results. The Shapley value for feature
i is defined as the weighted average of the marginal contributions of feature
i to all possible coalitions
S of features:
where
is the contribution of feature
i,
N is the set of all features,
S is a subset of
N not containing feature
i, and
is the value function, representing the model prediction for the subset
S. A linear function of binary features
is defined based on the following additive feature attribution method:
where
is the mean value of all the training samples;
M is the number of input features; and
. When a feature is observed,
; otherwise,
. We use the SHAP technique to evaluate the significance of each input feature and its contribution to the predicted result.
4.1.4. Performance Indices
In this study, the prediction results of the four models are evaluated using well-established statistical criteria. Specifically, these criteria include the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Coefficient of Determination (
), and Mean Squared Error (MSE). These metrics are crucial as they analyze the relationship between predicted and actual outputs. Higher
values, which is Approach 1, indicate a good correlation, while lower RMSE, MAE, and MSE values indicate higher model prediction performance. The mathematical formulas of these indices are presented in Equations (
A2)–(
A5) in
Appendix D.
4.2. Results
4.2.1. Scale Validation and Statistics Assessment
The statistic examinations are completed using R software (R 4.3.1 Version), and the packages used include
lavaan,
ltm, and
metan packages. The input variables are AVD, WOI, BOC, INNs, and SEPs, all derived from Sub-study 1. Additionally, common covariates such as years of work experience, educational level, age, and gender are considered. Previous studies have indicated that these factors may also influence employees’ boundary management strategies [
49,
50,
51]. Hence, socio-demographic characteristics are incorporated into the final predictive model. Unlike most machine learning-based predictive models that use directly measured variables as input features, this study employs a survey method for data collection. Therefore, a rigorous statistical examination is essential for enhancing prediction robustness. Following the data cleaning procedures mentioned earlier, we now proceed to a comprehensive statistical analysis.
The first step involves performing an exogenous model check. For each construct (e.g., AVD, WOI, etc.), we employ a series of items to measure the corresponding construct. We validate the scale to ensure that all items are suitable for measuring the construct and that the model fits the data well. The item loadings and Cronbach’s alpha are satisfactory (see
Table A1 in
Appendix B), indicating that the measurement used has sufficient reliability.
A Confirmatory Factor Analysis (CFA) is conducted using the lavaan package with the cfa() function in R (R 4.3.1 Version). The syntax is as follows:
mymodel <- cfa(model, data=mydata, std.lv=TRUE)
summary(mymodel)
The results demonstrate satisfactory fit indices, including CFI = 0.901, TLI = 0.887, RMSEA = 0.067 with a 90% CI [0.063, 0.071], and SRMR = 0.076. These results indicate that the survey instrument is both valid and reliable for measuring the latent constructs.
Second, we utilize plotting libraries in R to visualize the results from the Pearson coefficient. This analysis is executed using the
metan package. The mathematical expression of the Pearson correlation coefficient is shown in Equation (
A6) in
Appendix D. The result is presented in
Figure 3, where AVD represents Availability Demands, WOI denotes Work Identity, SEPs stands for Segmentation Preferences, BOC signifies Boundary Control, INNs indicates Integration Norms, and WWIN refers to Work–Non-Work Integration. The values in the upper diagonal are Pearson’s correlation coefficients. The target variable, WWIN, shows significant correlations with all independent variables. Additionally, the diagonal in
Figure 3 displays the frequency distributions of each variable, whereas the plots below the diagonal represent scatterplots showing relationships between variable pairs.
Third, for all covariates, the boxplot function is used to analyze their relationships with the target variable. As depicted in
Figure 4, employees from all age groups experience WWIN, with those aged between 30 and 40 years exhibiting higher median values of WWIN behavior. Moreover, employees with higher educational levels are more prone to experience increased WWIN. Employees with fewer than 5 years or more than 10 years of work experience are also more likely to encounter WWIN. There is no discernible relationship between gender and WWIN. The descriptive statistics of all variables are shown in
Table A2 in
Appendix C.
4.2.2. Model Selection
In ensemble learning techniques, hyperparameter optimization is crucial for enhancing prediction accuracy, especially in bagging-based approaches. Taking the XGBoost as an example, by manually specifying the key hyperparameters
and
, we are able to visualize the R-squared values in the contour plots in
Figure 5. It can be seen that the R-squared values improve when
is relegated to below 300, resulting in almost all R-squared values being above approximately 0.4, irrespective of variations in
. Nevertheless, this manual manipulation only displays the effects of two hyperparameters at a time and can be arduous. Instead, the K-Fold CV is more efficient for simultaneously tuning multiple hyperparameters. To this end, we use the aforementioned technique for hyperparameters optimization. The results of the model performance are shown in
Table 3.
It can be observed that both LightGBM and XGBoost models yield acceptable results, with values exceeding 0.5. XGBoost provides the highest value, coupled with the lowest RMSE, MSE, and MAE among all models considered in this study, thereby outperforming the other models. The performance metrics for the RF model are slightly inferior compared to those of LightGBM, while the LR model exhibits the poorest results. The average of the squares of the errors (MSE), which are the differences between the observed values and the values predicted by the model, is significantly reduced in the XGBoost model. These findings indicate that XGBoost holds significant advantages in forecasting WWIN behavior. The superior performance of machine learning models suggests that the relationships between the input features and WWIN are likely intricate and nonlinear.
To assess the predictive power of our analysis, we illustrate the top two superior models in
Figure 6, which depicts the prediction accuracy using R-squared values based on the 10% testing data. The orange line represents the predictive power of XGBoost, and it can be observed that the predictions are extremely close to the actual values in the test dataset. The effectiveness of the XGBoost algorithm has been validated in numerous existing studies across a wide range of subjects [
23,
52,
53,
54]. Our study corroborates these prior findings and further demonstrates its robust effectiveness for predicting human behavior in a managerial context.
4.2.3. SHAP Values Analysis
The SHAP interpretation analyses based on XGBoost prediction are depicted in
Figure 7 and
Figure 8. The SHAP summary plot (see
Figure 7) is a beeswarm plot displaying an information-dense summary of all samples, reflecting the distribution of the SHAP value of each feature devoted to the predicted results. The features in the Y-axis are arranged in ascending order according to the features’ importance. The X-axis position of a dot conveys information about the impact of the specific feature on the model output for that specific data point. The overlap point is jittered upwards on the Y-axis. The color shades correspond to the right-hand legend and reflect the feature value strength. According to
Figure 7, AVD is an essential feature, followed by SEPs, BOC, and INNs in determining the behavior prediction model. Specifically, higher AVD values result in higher SHAP values, corresponding to a high boundary integration behavior. On the contrary, lower AVD values could decrease such behavior. Interestingly, the distribution of SHAP values is asymmetrical along the perpendicular line between the negative and positive SHAP value regions relative to the X-axis. There is a ceiling effect for the positive impact of AVD on WWIN; once AVD reaches a certain threshold, its contribution to positive effects diminishes. However, on the opposite side, low AVD values can continuously reduce the SHAP values significantly away from the perpendicular line. Furthermore, for moderate levels of AVD (dots colored in purple), the dots are predominantly located to the left of the SHAP value equal to zero, suggesting that a moderate level of AVD exerts minimal positive effects on WWIN. In a nutshell, AVD is the most important feature in predicting WWIN behavior and its impacts are nonlinear.
We move on to interpret the importance of the remaining features. Features SEPs and BOC generally show similar patterns in prediction. Features SEPs and BOC reveal similar patterns in prediction. Specifically, higher values of SEPs and BOC are associated with lower SHAP values, indicating lower integration behavior, and vice versa. The difference between SEPs and BOC becomes evident when considering individuals with moderate levels of these predictors. For SEPs, dots are primarily located to the right of the perpendicular line compared to BOC, suggesting that people with a moderate level of SEPs are more inclined to integrate their work and life, reflected by positive SHAP values. In other words, individuals with a moderate level of boundary control tend to implement stricter strategies to maintain a separation between work and life, resulting in low or nearly zero integration predictions. Conversely, those with moderate segmentation preferences (i.e., a tolerance for a permeable boundary) are more inclined to engage in boundary integration behavior. This finding helps us differentiate the roles of SEPs and BOC, which are often used interchangeably. Regarding INNs and WOI, their impacts on WWIN are minimal. For very high values of INNs and WOI, positive SHAP values are observed. However, in all other cases, their impacts on WWIN are generally negative. With respect to socio-demographic characteristics, the higher the values of working years and educational level, the more likely it is that an integration behavior occurs, which is consistent with intuitive expectations. For other characteristics, the effects are miscellaneous with no clear patterns observed.
The contribution of a feature often depends on the presence of another feature. Therefore, we draw double feature dependence plots to illustrate the relationships between features. Given that AVD is the most important feature, we calculate the top two relevant features associated with it. The results are shown in
Figure 8. The X-axis indicates the value of the main feature, while the Y-axis represents the SHAP value of this feature in each data sample. The color bars to the right of each subplot display the values of the variable most dependent on the feature represented by the X-axis. We observe that the most important features interacting with AVD are INNs, followed by BOC. A common attribute in both plots is that when the value of AVD is less than four, the distribution of SHAP values spans a large range. Higher AVD corresponds to higher SHAP values, indicating increased positive effects for individuals experiencing higher injunctive norms. Conversely, a higher AVD combined with lower BOC increases the likelihood of integration behavior.
5. Discussion
Theoretically, the results address the research question of whether mobile platforms are the presumptive culprits of work–life integration. We analyzed the nonlinear behavioral patterns associated with technology-induced work–life imbalance. From a methodological perspective, we developed a fully data-driven approach that includes the co-occurrence network technique and an explainable AI framework incorporating machine learning algorithms and SHAP techniques. Our discussions were organized around the theoretical, practical, and methodological contributions as follows.
First, we found that the availability demands triggered by mobile platforms are the most influential factor leading employees to tolerate the encroachment of work into their non-work domains. The convenience and connectivity provided by these platforms afford employees greater flexibility but also result in prolonged working hours. This aligns with critiques that highlight the darker aspects of technology, describing it as a tyrant (see article [
55]). Our results, grounded in an explainable AI framework, suggest that individuals experiencing low availability demands are more inclined to maintain a clear separation between their work and personal lives. Conversely, increased availability demands on mobile platforms do not necessarily correlate with a continuous tolerance for a permeable and integrated work–life boundary. This finding extends prior studies by demonstrating that there is a threshold for the positive association between mobile platform usage and work–life integration behavior, suggesting that the negative aspects of technology might be exaggerated [
56]. While availability demands facilitated by mobile platforms enable employees to work anywhere and anytime, as argued in [
57], the decision to integrate work into personal life largely remains at their discretion.
Second, in juxtaposing the nonlinear relationships observed, we identified two behavioral patterns—work–life integration and separation—underscoring the importance of considering the dichotomy of such behaviors. To enhance work–life integration, executives and managers should recognize that merely imposing high availability demands does not necessarily result in employees being more diligent or productive. Other factors, such as integration norms and high job autonomy, are more significant in influencing employees to work during non-work hours. In other words, whether employees choose to work during their leisure time is a complex decision influenced by their colleagues’ behaviors and the degree of job autonomy they possess.
Conversely, to mitigate any imbalance and promote sustainable human resource development, organizations could adopt strategies to discourage integration behavior. This can be achieved by reducing availability demands on mobile platforms and fostering an environment that emphasizes the separation of work and personal life. In this regard, our study diverges from prior research that predominantly addresses a single perspective, thereby failing to provide a comprehensive solution for organizations.
By highlighting the multifaceted nature of work–life integration and separation, we offer a holistic understanding of the impacts of mobile platform usage. This approach allows organizations to tailor their strategies effectively and agilely according to their specific targets, balancing the benefits and drawbacks of mobile technology platforms with their needs.
Methodologically, our contributions are twofold. Instead of relying on a contrived scheme to manually sort out all determinants of integration behavior, we utilized a robust data-driven method. Specifically, we innovatively identified key phrases from the existing literature using qualitative analysis tools and subsequently applied a co-occurrence network to calculate the nodes and links between variables. This approach allowed us to explore less apparent determinants that might have been overlooked due to the constraints of specific theoretical frameworks. During the feature extraction process, we observed that much of the existing literature on this topic selects variables based on instinct or a theoretical frame, which is prone to preference bias. By extracting influencing factors grounded in the existing knowledge, we were able to unravel a comprehensive scene graph of integration behavior.
In addition, we incorporated principles from behavioral sciences into an explainable AI framework. While a vast body of existing studies employs numerical input features that are directly measured (e.g., [
38,
54]), our approach diverged by using survey data to build a predictive model. Through rigorous statistical examination, we demonstrated that using survey data for predictive modeling is both feasible and effective. This responds to the recent call for expanding the application of AI in various managerial studies [
58], which are often intertwined with survey data and linear regression models.
6. Conclusions
In conclusion, by developing a completely data-driven methodology, our findings can be summarized as follows. First, we identified that beyond the availability demands on mobile platforms, other factors such as segmentation preferences, boundary control, integration norms, and work identity also play important roles in shaping work–life integration behavior. The results from the explainable AI framework indicate that availability demands are the most significant contributing factor, yet their positive and negative impacts are not equally distributed.
In instances where availability demands are extremely high, employees tend to take matters into their own hands and do not allow work to intrude upon their personal lives without control. Specifically, they may tolerate overtime work caused by mobile platforms, but the positive influences of such tolerance are neither as considerable nor as persistent as assumed in linear regression models.
Furthermore, individuals who prioritize their work identity more prominently tend to engage in work–life integration when faced with higher availability demands. In this case, their actions align with their role identity and are likely to be self-motivated. Conversely, employees with limited job autonomy and control are also likely to adopt work–life integration behaviors under high availability demands. However, their actions are more likely to be organizationally enforced and determined by job characteristics, with little individual preference.
These nuanced insights highlight the complexity of work–life integration behaviors and the need for personalized strategies in managing work and personal life boundaries in a mobile technology-driven workplace.
Our study makes a significant contribution to the academic discourse by addressing the nonlinear and multifaceted nature of work–life integration behaviors influenced by mobile platforms. We offer a nuanced perspective that challenges existing linear models, thereby enriching the scholarship on technology-induced work–life dynamics. Orienting the AI-empowered workplace towards a sustainable future requires a thorough understanding of how technology revolutionizes our daily activities. Such understanding aids scholars and practitioners in formulating strategies for a sustainable future, considering the intricate interactions between humans and technology.
Our methodological contributions are also significant. The introduction of a fully data-driven methodology, featuring the co-occurrence network technique and an explainable AI framework with machine learning algorithms and SHAP techniques, provides a robust and transparent way to understand complex behavior patterns. This methodology not only enhances the accuracy and validity of our findings but also offers a replicable framework for future research. The use of explainable AI techniques ensures that the results are interpretable and actionable, thus bridging the gap between theoretical insights and practical applications.
Our study does, however, have some limitations. It is important to note that SHAP values provide interpretations based on associations, akin to linear regression models, and do not establish causality. Our analyses are susceptible to omitted variable bias and do not encompass all socio-demographic variables due to ethical regulations and anonymity considerations. For instance, we did not consider indicators such as leader support or self-efficacy, which may affect human–technology interaction. Additionally, reaching a representative sample would be conducive to building a robust predictive model. Moreover, we relied on only four analytical methods. Future research could benefit from applying a broader range of models, such as Artificial Neural Networks (ANNs), Support Vector Machines (SVMs), and hybrid machine learning models, to gain a deeper understanding of the phenomenon. Nonetheless, this study represents a significant step forward in developing interpretable machine learning models to explore complex relationships, thereby broadening the potential application of ML algorithms in managerial contexts.