Computing grounded theory: a quantitative method to develop theories

,

Page 2 of 26 Chen and Chen The Journal of Chinese Sociology (2024) 11:17 philosophical speculation that does not involve social phenomena; however, it seems to have gradually lost the initiative in theoretical production and development in long-term competition with qualitative research.Qualitative researchers constantly observe, discover, and refine new concepts and theories, forming the pioneer of theoretical development.In contrast, quantitative researchers conduct postpositional statistical tests on theories or hypotheses based on literature and the sociological imagination.Quantitative researchers value Karl Popper's definition of science, and, consequently, often become lost in their self-appreciation of the importance of falsification testing, unconsciously neglecting the value of data and models for directly inspiring theory and the potential application of inductive logic in quantitative research.
Is there a new logical approach and model that allows quantitative scholars not only to conduct postpositional scientific tests but also to directly generate hypothetical theory with data?In fact, there is a long history of developing theoretical explanations with quantitative data, such as in Émile Durkheim's classic study on suicide.There are various feature selection methods used in statistics to meet that end.However, quantitative data is infrequently used to develop theory after decades of institutionalization of quantitative research methods.This is probably because random or even ergodic brainstorming of N × (N − 1) pairwise correlations over the variable list may generate a large number of unfounded or even absurd hypotheses, and loop testing of various X-Y combinations with traditional regression models cannot solve many problems, such as limiting the number of control variables, selecting combinations, and multicollinearity.Therefore, the direct generation of theory with data and models has been neglected by quantitative scholars for a considerable period.
However, it is possible at present with the increase in large-scale social survey data and the application of machine learning in sociology.In this article, we propose a quantitative theoretical production method based on a large amount of data and machine learning models: for a given Y and a large number of explanatory variables X, the predictive ability of X for Y is quantitatively analyzed with a supervised learning method.By utilizing the logical relationship between causality and predictability, we can explore and screen for many X with strong predictive power, directly leading to the generation of theory, finding potential new X with theoretical value for Y, and helping sociologists generate, develop, and revise theories.Although this method is a typical computational social science method, its logical starting point shares similarities with the core principle of grounded theory.It goes beyond the preconceptions of theory and digs into the data itself without making any theoretical hypotheses, thereby transcending the logic of deduction verification and conducting theoretical research with empirical cases.Therefore, we call this "Computing Grounded Theory". 1his article will first briefly present the hypothesis testing methods of traditional quantitative research and then provide a detailed introduction to the specific logic and approach of Computing Grounded Theory.The article then further examines the possibility of Computing Grounded Theory from both theoretical and methodological perspectives and uses the sense of well-being as an example to demonstrate the reliability and validity of Computing Grounded Theory.Finally, this article concludes with a critical review of the methodological significance and potential problems that may arise from Computing Grounded Theory.

Hypothesis testing and the wheel of science
In the past four decades, the mainstream pattern of quantitative sociology research has been based on survey questionnaire data and the use of multiple regression models to statistically infer whether there is a correlation or causality between the explanatory and dependent variables.Yusheng Peng once vividly compared quantitative research to "Westernized eight-legged essays" (Peng 2010:180).He noted that mainstream social science journals adopt a relatively standardized "template style" format, with each part of eight sections performing its own duties interconnectedly-question, literature, hypothesis, measurement, data, method, analysis, and conclusion (Peng 2010).Scholars from other countries have made similar summaries of hypothesis testing models by analyzing the content of the American Sociological Review (Wells and Picou 1981).More interestingly, the hypothesis-testing paradigm of quantitative research is not unique to sociology, which has permeated various disciplines, such as economics, political science, and psychology (Lin 1995).Although the relevant parts can be merged or refined, the basic principle is to falsify the proposed null hypothesis.
However, testing theory is not the entire work of scientific research.Walter Wallace proposed the wheel of science (Fig. 1) in The logic of science in sociology, pointing out that sociological research is a cyclic, spiral, and endless process that includes theory construction and testing (Wallace 1971).It is obvious that the quantitative paradigm of hypothesis testing is confined to the right half of the scientific wheel.It was originally a complete path of scientific research from theory construction to theory testing, but with the distinction between quantitative and qualitative research, theory construction seems to have become the exclusive mission of qualitative research, and quantitative research increasingly adheres to theoretical verification as the norm.In fact, as Merton suggested, empirical research goes far beyond the passive function of testing theories.It not only confirms or refutes hypotheses but also performs at least four functions in the development of theories: creating, revising, transforming, and clarifying theories (Merton 2006)

The historical origin of hypothesis testing
The hypothesis testing paradigm originates from the positivist tradition and was strengthened after the standardized movement of quantitative research by the Columbian school.Paul Lazarsfeld and colleagues (1967) and Samuel Stouffer (1962) advocated for the purpose orientation and scientific movement of using empirical materials to validate theories.Stouffer wrote the book Social Research to Test Ideas, which further popularized the method of using data to verify theory throughout the field of quantitative research.This methodological tradition does not regard traditional theoretical discourse, which contains a large amount of metaphysical speculation and untested assertions, as precise scientific knowledge, since it is just empty statements that do not improve reliable judgments about social facts.Through academic cultivation and methodological training, quantitative sociological researchers have gradually developed an empirical personality that requires them to constantly revise their ideas about society and commit to improving the effectiveness of social science by answering substantive questions (Pawson 2000).
There is a consensus in sociology that the quantitative paradigm of hypothesis testing bridges the gap between theory and experience, ensuring the scientificity of conclusions.However, if we took the analytical approach of theory first-data validation next as granted without reflection, this methodology, which used to be a force for knowledge liberation, would easily be transformed into constraints on the productive creativity of theory.In fact, using quantitative data for theory exploration is not new.As early as forty years ago, analytical methods and models for automatically selecting variables from data were available.Statistically, methods such as forward selection, backward selection, and stepwise regression are used to select the most suitable variables for the model.Subsequently, partial least squares regression based on feature dimension reduction, AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) standard comparisons based on information criteria, and regularization methods such as ridge regression and LASSO (Least absolute shrinkage and selection operator) regression have emerged.Although these methods have been somewhat criticized either for unscientific variable selection criteria or for the equivalence between correlation and causality (Rubin 1974), quantitative scholars have been inspired by the data itself in the actual research process.However, many researchers do not present their research according to the actual research process after obtaining new discoveries from the data; instead, they then tends to write their induction discovered hypothesis as if it had been thought up before the research began, and then prove it according to the logic of hypothesis testing (Glaser 2008;Wu and Li 2020).

Consequences of hypothesis testing only
The hypothesis testing approach of quantitative research constrains the knowledge production of quantitative research in two ways.

A lack of quantitative exploratory research
The academic community has gradually formed a common impression among practitioners and bystanders for a long time that empirical research is used to verify theories, and theories are obtained through researchers' whims (Merton and Barber 2011).Although it has contributed to the common prosperity of different research paradigms, it has also led to the postposition or even absence of quantitative research in the field of scientific discovery: quantitative research has refined existing theories but has rarely produced new theoretical constructions (Charmaz 2009).

Excessive reliance on common sense and the absence of insight in quantitative research
The hypotheses validated by quantitative research are primarily derived from existing theories or sociologists' common sense and inspiration.The contradiction of common sense lies in its ability to help us understand the world while also weakening our ability to understand (Watts 2011).Interestingly, on the one hand, sociologists need to doubt and verify the scientific nature of common sense with a disciplinary mission of challenging common sense; on the other hand, they have to select possible explanatory variables from existing common-sense stereotypes when establishing hypotheses, which often leads to doubts -using complex methods to verify common sense (Liu and Gong 2020:155).

Computing grounded theory: theory generation with machine learning
The fundamental idea of computational is to bridge a reverse path from data to theory, utilizing the predictive power of machine learning and interpretable attribution algorithms and directly generating the mechanism theory of established dependent variables based on the law that causality is a sufficient and unnecessary condition for predictability.This section will provide a detailed discussion of the basic ideas, logical premises, and methodological foundations of Computing Grounded Theory.

The basic ideas of Computing Grounded Theory
As shown in Fig. 2, Computing Grounded Theory includes the following basic six steps.
The first step is to set the dependent variable.Based on the data from the social survey questionnaire, research Subject Y was selected based on research interests and needs.Theoretically, we cannot determine Y beforehand; thus, each non-preassigned variable becomes the predicted object Y and is analyzed with the ergodic exploratory method.
The second step is to prepare high-dimensional data.Social survey data are often highdimensional, with hundreds or even more variables.Each of these numerous indicators may be a potential cause of Y, which implies the possibility of grounded theory.Data at different dimensions can be matched, and even irrelevant features can be included.
The third step is to carry out social prediction.Based on high-dimensional data, supervised learning methods, such as support vector machines, random forests, gradient boosting trees, and neural networks, are used to train the prediction model of Y. Algorithms can be diverse.As long as relatively good prediction results can be achieved, complex or interpretable algorithms could also be considered.
The fourth step is to compare the predictive performance.With the interpretability algorithms of machine learning models, attribution analysis is performed on the black box models generated by predictions, and possible causes are searched for based on the ranking of X's predictive power on Y.The basic idea is to determine whether disrupting certain features affects the accuracy of model prediction and how changing features will affect the prediction results.
The fifth step is to search for potential theories.Social associations not addressed in previous studies are searched for based on a set of Xs sorted by predictive power.They can be compared with existing research to validate or clarify theories.Similar explanatory terms can also be classified as abstract concepts or inductive theoretical propositions.
The sixth step is cross-validation to verify the robustness of the results of Computing Grounded Theory and the applicability of the theoretical hypotheses.Different data, other machine learning models and attribution algorithms are adopted to calculate the same dependent variable.Alternatively, other hypotheses derived from the new theory are reexamined to mutually verify and complete the wheel of science.
In brief, this method is vastly different from quantitative research on hypothesis testing.It does not presuppose theoretical hypotheses but relies solely on algorithms and data to train predictive models.By accurately estimating and ranking the predictive power of X on Y, it provides a set of potential theoretical hypotheses for possible causal relationships, ensuring an indiscriminate open attitude to data and precise goal orientation toward empirical problems.According to the four-quadrant framework of social science research methods proposed by Hofman and co-authors (2021), grounded computational theory is a comprehensive modeling approach that integrates prediction and interpretation.According to the five major types of social science methods classified by Zhou and colleagues (2022), the basic ideas for Computing Grounded Theory are "exploratory research based on (big) data" (Zhou et al. 2022:133).If other data are used to further validate the exploratory theoretical hypothesis after it is generated, it belongs to the "integrated research of exploration followed by verification" (Zhou et al. 2022:143).
In fact, scientific research assisted by machine learning is currently a cutting-edge research trend.Nature once predicted that artificial intelligence would guide scientific intuition under the cover title of "AI-guided intuition".Specific paths for using similar methods to guide intuition and propose guessing have also emerged in the fields of mathematics (Davies et al. 2021), economics (Ludwig & Mullainathan 2024), and management (Shrestha et al. 2021 2022) have proposed a similar analytical strategy for team innovation ability.On the basis of these relevant empirical explorations, we could standardize and systematically refine this approach and propose a complete method and practice approach that hinges on the possibility of its application in sociological research, focusing on its universal methodological value and significant differences from traditional research models, and substantively connecting the interpretability of algorithmic models and theoretical generation algorithmic models at the methodological level.

The logical premises of computing grounded theory
As a method of theory generation, Computing Grounded Theory has clear logical premises, comprising two aspects.
One is the inductive logic of grounded theory.Grounded theory arises from the reflection of empirical research on the paradigm of quantitative hypothesis testing.Grounded theory's founders noted that sociology places too much emphasis on theoretical verification, and "attempts to close the gap between theory and research have concentrated principally on the improvement of methods for testing theory" (Glaser and Strauss 1967: vii).Researchers should discover theories from data to bridge the discrepancy between empirical research and theoretical research.Grounded theory advocates extracting theories directly from empirical materials with a stepwise induction method and comparing them with existing theories and research.Avoiding preconceived ideas or speculations before the analysis is an important principle to ensure the effectiveness of grounding.
Notably, Barney Glaser, as the founder, emphasized that grounded theory is a universal methodology that applies to both qualitative and quantitative data."There is no fundamental clash between the purposes and capacities of qualitative and quantitative methods or data, "... "each form of data is useful for both verification and generation of theory" (Glaser and Strauss, 1967: 17-18).However, with the development of grounded theory, people have found that it still seems more suitable for qualitative research.Anselm Strauss, another proposer of grounded theory, even regarded it as an exclusive tool for qualitative research (Strauss and Corbin 1994).The reason is not difficult to comprehend: the depth and interpretability of qualitative data are often more conducive to directly proposing theoretical hypotheses with sociological imagination, while quantitative data, as a numerical indicator, are characterized by highly simplified abstraction, and their inherent mathematical and statistical correlations are difficult to discover through intuitive means.In fact, to overcome the stereotype in the academic community that grounded theory is only applicable to qualitative data, Glaser has written a manual titled "Doing Quantitative Grounded Theory" to elaborate upon the steps of quantitative grounded theory.The basic idea is: "Saturating core index with all possible two-variable runs, discovering relationships among the theoretically relevant consistency indices, summation indices and single questionnaire items, and then generating conceptual hypotheses.The next non-neglectable step is elaboration analysis.That is to make three or more variable analyses in order to saturate categories further by developing their properties and thereby achieving a denser theory".(Glaser 2008: 54).However, the problem is that a large number of variables are difficult to directly correlate with the human brain.When using statistical methods, there is actually a lack of clear selection criteria for which variables should be included in the model.In particular, when the number of independent variables is large, there may be many problems, such as insufficient degrees of freedom or collinearity.In short, the logic of quantitative grounded theory is feasible, but there is currently no suitable method to carry out a convincing application.
The other aspect is the predictable logic of causality.The predictability and causal mechanisms between social phenomena are two different but highly correlated categories.According to Max Weber, sociology is a science that provides causal explanations of behavioral processes and outcomes (Weber 1968), in which sociological theories can be understood as the causal relationship between indicators.According to this logic, the dependent variable in sociological indicators must have predictability for the independent variable.This is because predictability is a necessary but insufficient condition for establishing causal relationships.It is also the most basic means to verify the principle of a mechanism (Watts 2014).
However, due to limitations in mathematical and statistical tools, sociologists often do not pay much attention to predictions.When discussing the concepts of cause and effect, correlation, and prediction in sociology, sociologists often use evasions: sometimes they emphasize that prediction is not equal to cause and effect, while abandoning the logic that cause and effect can inevitably be predicted.Alternatively, they argue that complex regression models that incorporate too many independent variables are not concise enough, or they criticize algorithmic models that can make data predictions unexplainable due to black box processes.Duncan Watts has summarized and strongly criticized these types of arguments (Watts 2014).
One of the logical foundations of Computing Grounded Theory is to fully utilize the important relationship between prediction and causality; that is, causality is a sufficient but unnecessary condition for prediction.This means that if an X can predict Y well, X may indeed cause Y.Although this relationship is only possible, not inevitable, its probability of forming causality is much greater than that of nonpredictive associations.In the context of sociologists gradually confining the focus of their discipline to two-variable analysis and abandoning social prediction (Hofman et al. 2017), it is important for quantitative research to generate theories with the predictive power of machine learning.

The implementation of computing grounded theory
Computing Grounded Theory allows for the interaction of dozens, hundreds, and even thousands of variables, which is much more stable and reliable in theory discovery by comparing the predictive power of the relevant eigenvalues of algorithmic models than through human thinking.The specific implementation process includes two aspects: social prediction and predictive power comparison.

Social prediction: fitting algorithms with supervised learning
Traditional quantitative regression models are good at correlation and causal inference rather than prediction.Therefore, what kind of model is most suitable for predicting complex social processes?Leo Breiman, a renowned statistician, divided statistical modeling methods into two groups: data models and algorithmic models.The data model assumes that the data follow a certain functional distribution f (x) in advance (such as a linear regression model) and then fits and estimates the parameters of the assumed f (x).The algorithmic model does not assume any distribution characteristics of the data and aims to find a function g (x) through which y can be predicted (Breiman 2001a).In fact, this classification precisely highlights the fundamental difference between traditional econometric models in sociology and machine learning.Breiman further noted that the mindset of data models that are widely used in social and behavioral sciences emphasizes unbiased estimation of model parameters rather than predictive accuracy.In other words, the widely recognized practice model in social science is not to inquire whether specific data and models can predict certain interesting results but to inquire whether specific coefficients in idealized models are statistically significant and their impact direction.
However, there are two obvious problems with data models.First, the data must meet certain assumptions to fit specific parameter models.Taking linear regression as an example, the relationship between independent and dependent variables needs to be linear, the respective variables are not multicollinear, the residuals should follow a normal distribution, the perturbation term should meet the same variance, and there should be no autocorrelation.However, in complex and diverse societies, certain data are too strict.Therefore, the academic community has adopted an ostrich policy, gradually shifting its focus to statistical significance and maintaining an open or suspenseful attitude toward whether the data meet model assumptions (Freedman 1991).Second, the conclusion is about the mechanism of the model rather than the mechanism of the facts.Imposing simple parameter models on the data generated by complex systems can lead to a loss of accurate and critical information.Model errors or the introduction of a large amount of discretion by researchers in data analysis can result in potential biases (Simmons et al. 2011).If the model cannot simulate natural situations, the conclusion would be incorrect (Breiman 2001a).
The algorithmic model, represented by machine learning, provides a rather powerful alternative solution for the above problems.The implicit epistemological assumption of the algorithmic model is that the intrinsic mechanism of factual data is unknown and complex, and the key is to find an algorithm that can predict y well with x, that is, to use the algorithm to fit the data.Algorithmic models often adopt nonlinear, nonparametric estimation methods to adjust the complexity of the model through one or more hyperparameters.Including data complexity in machine learning provides the analyzed data with arbitrary distributions without any assumptions.We suggest that this liberation will advance theory generation in at least two ways.
First, it reflects the nonlinear data relationships that exist in real social processes.The linear assumption of data models often hardly reflects the social reality.Although the model's simplicity can be advantageous, simplification is only a means rather than an end.Most machine learning fitting processes do not have to fulfill existing function settings but rather aim to pursue prediction accuracy as the highest objective (Breiman 2001a).
Second, it reflects high-dimensional complex data relationships in real social processes.Traditional econometric models can incorporate only limited explanatory variables.Supervised learning algorithms can simultaneously consider thousands of different factors and various complex interaction patterns in a single learning model (Linthicum et al. 2019).The influencing factors of a social phenomenon are numerous and complex, and incorporating more potential 'causes' increases the likelihood of discovering new explanatory dimensions.

Comparison of predictive power: attribution algorithms for the interpretability of black box models
Although machine learning has challenged the various limitations of existing statistical models and improved the simulation of the true state of things, it is widely criticized for its black box process, resulting in inexplicability.However, an increasing amount of evidence in recent machine learning literature suggests that the contradiction between predictive accuracy and interpretability is not as severe as imagined.With the urgent demand for the interpretability of complex models, an increasing number of methods for disassembling black boxes have been invented and implemented (Ribeiro et al. 2016).A widely cited paper by Harvard professors noted that the explanatory analysis of the black box model in machine learning is an effective method for revealing interpretable factors based on data (Doshi-Velez and Kim 2017).
The SHapley Additive exPlanations (SHAP) method could be an example of a detailed explanation of the specific mechanism of the black box model.This method calculates the Shapley value of each X based on alliance game theory and considers it an indicator of its significance.Given that the number and sequence of different participants affect the final overall return, this method calculates the difference in overall return among various states, such as including and excluding the participants, for each combination and records it as the marginal contribution of the individual participant by exhausting the arrangement and combination of various participants.Then, it calculates the mean of the marginal contribution of the participant for various arrangements and combinations and records it as the Shapley value of the participant (Shapley 1953).The sum of the Shapley values of all participants represents the overall return.
The specific calculation formula for the Shapley value of each participant i is: N is the set composed of all participants.|N| is the number of participants included in this set.S is a combination of participants and a subset of N. S ⊆ N\{i} represents all subsets of set N after removing participant i. v(S ∪ {i}) − v(S) is the average marginal contribution of the overall return with participant i to the total return without participant i, which appears | S |! (| N | − | S | − 1)! in the total ranking.Data scientists have developed the SHAP algorithm to improve computational efficiency, which approximates the Shapley value through conditional expectation functions.The specific techniques will not be detailed here.The Shapley value fully considers the interaction between variables and has a solid foundation in game theory.It is a fair allocation method with simultaneous effectiveness, symmetry, virtuality and additivity (Lundberg and Lee 2017).
In addition to the Shapley value, there are many other methods for interpretability analysis of black box models, such as permutation feature importance, which measures the significance of a particular feature by comparing the changes in model prediction errors before and after permutation (Breiman 2001b).Another example is the partial dependence plot, which is a graphical representation that shows how a single feature influences the predicted output of a machine learning model while controlling for the effects of other features (Zhao and Trevor 2021).Alternatively, an interpretable local surrogate model can be used to simulate the original black box model (Ribeiro et al. 2016).Establishing and developing these methods have provided the possibility of rebalancing the accuracy and comprehensibility of predictions, laying a solid methodological foundation for Computing Grounded Theory.

Research question and data
We use subjective well-being to demonstrate how Computing Grounded Theory could help to inspire and clarify the theory of well-being.The data used in this case are extracted from the Chinese General Social Survey (CGSS) of 2017, which includes a total of 12,582 samples and 783 variables, providing a relatively comprehensive and reliable dataset for the calculation and analysis of well-being.The dependent variable of this study is "Overall, do you think life is happy -very unhappy, relatively unhappy, hard to say, relatively happy, or very happy?", and the independent variables are all variables in the questionnaire except for the dependent variable.3

Research methods and steps
The first step is data preprocessing.First, a binary Y helps improve the accuracy of the algorithm's prediction.We mark "very unhappy, relatively unhappy, and hard to say" as 0, representing the unhappy samples, and record the responses of "relatively happy" and "very happy" as 1, representing the happy samples.Second, we convert the categorical variables into dummy variables.Next, we delete variables with more than 30% missing values.Finally, since the number of class 1 samples is significantly greater than the number of class 0 samples, data imbalance may lead to algorithm bias.We use bootstrap sampling to oversample the minority class and achieve rebalancing between the two classes.
The second step is model training.We used the gradient boosting algorithm XGBoost to train the prediction model with 1000 subdecision trees and other default parameters.After 70% of the training set converges iteratively, the remaining 30% of the test set shows a model accuracy of 0.92, with a recall rate of 0.86 and an F1 score of 0.92.The overall performance of the model is satisfactory.
The third step is model attribution.The analysis uses SHAP, a model-agnostic method for interpreting any machine learning model, to identify the most influential factors and their impact on predictions.Specifically, we calculate a SHAP value for each independent variable X in each case.The connotation of this indicator is as follows: for this case, how much of an average marginal contribution will it make to the predicted results when this X is added compared to when it is not added?A positive value means that the addition of X leads to an increase in well-being, while a negative value means that the addition of X leads to a decrease in well-being.

Primary findings
Figure 3a shows the top 20 variables extracted by the attribution algorithm that contribute the most to predicting well-being, measured by the average absolute value of the SHAP value on each X for all cases, that is, the average marginal contribution of variable X. Figure 3b shows the details of the influence of different predictive variables in the form of scatter plots.Each point in the graph represents a real sample.For each row, the color represents the magnitude of feature X of the variable in that row.The darker the color of the point, the larger the X. 4 The horizontal axis represents the size of the SHAP value.The more points with the same SHAP value, the larger the cross-sectional area of the honeycomb, and the thicker it appears.Overall, the graph can reflect the way and magnitude of interactions among variables, as well as the distribution of individual cases.Taking the sense of fairness as an example, the scatter plot shows that cases with a greater sense of fairness (black dots) often concentrate on the right side of the horizontal axis; that is, a positive SHAP reflects an increase in well-being, while cases with a lower sense of fairness (gray dots) often concentrate on the left side of the horizontal axis, where the SHAP is negative and well-being is thus reduced.This indicates that a sense of fairness has a typical positive impact on well-being.
Figure 3 shows the top 20 variables that have the greatest impact on predicting happiness.The most significant predictor of wellbeing is the sense of fairness, with a SHAP value of approximately 1.4, followed by the level of health.For the convenience of induction, we categorize five primary influencing dimensions of wellbeing based on the similarity of variable meanings: (1) subjective cognition: sense of fairness and trust; (2) subjective and objective status: self-positioning of class, self-expectations of class after 10 years, family positioning of class, self-evaluation of socioeconomic status, and housing area; (3) demographic and health factors: level of health, year of birth, and health impact; (4) marriage and family: coresidential spouse, spouse's weekly working hours, spouse's annual income, and division of chores between husband and wife; and (5) lifestyle: rest and relaxation, watching TV, listening to music, and weekly working hours.Legend: In 3a, the average absolute value of the SHAP value for each variable (X) across all cases is depicted, representing the average marginal contribution of each variable.Figure 3b shows the details of the influence of different predictive variables in the form of scatter plots.Each point in Fig. 3b represents a real sample.For each row, the color represents the magnitude of feature X of the variable in that row.The larger the X is, the darker the color of the point.The horizontal axis represents the size of the SHAP value.The more points with the same SHAP value, the larger the cross-sectional area of the honeycomb and the thicker it appears.Taking the sense of fairness as an example, the scatter plot shows that cases with a greater sense of fairness (black dots) often concentrate on the right side of the horizontal axis; that is, a positive SHAP reflects an increase in happiness, while cases with a lower sense of fairness (gray dots) often concentrate on the left side of the horizontal axis, where the SHAP is negative and happiness is thus reduced.This indicates that a sense of fairness has a typical positive impact on happiness Theoretically, we can generalize all categories by layer and extract higher-level concepts and an overall theoretical model of wellbeing.It is also possible to further explore and compare a certain variable or specific dimension that has not been previously studied, explore common factors and covariate laws, and summarize theoretical hypotheses at the micro level.Given that the above variables and dimensions involve multiple disciplinary fields and have been discussed in previous theoretical and empirical studies (Liu et al. 2012;Qiu and Li 2012;Diener et al. 2018), we prioritize predictive power and select the variable "spouse's weekly working hours", which ranks among the top ten in terms of predictive power but has not been previously studied, for demonstration.

New discovery of wellbeing: searching for new variables with strong predictive power
The process of generating theoretical hypotheses from Computing Grounded Theory is composed of the following steps: (1) generating empirical propositions of potential hypotheses; (2) eliminating false correlations and establishing causal relationships; (3) inducing and conceptual refinement of relevant categories; (4) discussing existing theories and logical derivations; and (5) summarizing theoretical propositions and using other data methods for revalidation.The first step is to directly rank the predictive power of variables and discover factual propositions about the correlation between variables.However, the proposition has not yet established a rational understanding based on causality between phenomenon and essence.We can further deploy steps 2-5 to fill the gap between proposition and theory to increase the theory's scientificity. 5e first propose the empirical proposition.According to Fig. 3a, the variable "spouse's weekly working hours" ranks 9th in prediction, but previous studies have not paid sufficient attention to this variable.We present the relationship between the two variables as a new empirical proposition: a spouse's working hours can affect the other spouse's subjective well-being.
The second step is to adopt double machine learning (Chernozhukov et al. 2018) to exclude other possible confounding variables as much as possible, purifying the relationship between the two variables.Using all other questionnaire variables as confounding variables, four algorithms-Lasso, random forest, decision tree, and XGBoost-all showed a significant causal relationship between the two variables.Due to the word limit of this article, detailed results are not presented here.
The third step is to seek other variables that are highly similar to the connotation of X, observe whether they have explanatory stability and logicality, and then create a concept or a set of concepts to induce a unified understanding of the relationship patterns among the data to eliminate the predictive power caused by data randomness.In this example, "a spouse's weekly working hours" refers to a spouse's time allocation between work and family.We screen other similar variables with large SHAP values that indicate the time allocation of husband and wife to work and family: "weekly working hours (ranked 19)' and 'face-to-face communication time in the family (ranked 21)".
We further compare the above three variables to generate theoretical intuition.Figure 4 shows the variation curve of the SHAP values of the three variables through the partial dependence plots of the entire sample.It shows that the weekly self-working hours of both spouses increase their wellbeing within 0 to 40 h as working hours increase.However, after working for more than 40 h, there is a completely opposite result of the working hours of both spouses: longer working hours by spouses can significantly improve wellbeing, but an increase in self-working hours can significantly reduce wellbeing.This implies that if spouses take on more social responsibilities for the family and reduce the amount of time they spend at home, this would improve their partner's wellbeing.We conceptualize this phenomenon as "family distance".Moreover, excessive family distance, which means that spouses work more than 60 h per week, will weaken the positive influence of family distance on people's wellbeing.Similarly, the "face-toface communication time in the family" also shows a peak distribution, where the best length of time is approximately 20 h per week, and the positive influence of the communication is relatively weakened when family members communicate with each other for less than or more than 20 h.By comparing the above variables, we find that improving well-being requires a certain family distance between spouses, but this distance should not be too close or too far.Thus, we construct the core theoretical hypothesis that family distance affects the wellbeing through variable comparison and conceptualization.
For step four, we need to engage in a dialogue with existing theories and logically derive detailed mechanisms of causal relationships between variables, including influence pathways (mediating effects) and heterogeneity of different group influences (moderating effects), to form an abundant series of logically progressive hypothesis propositions. 6sychological research suggests that the allocation of time, individual autonomy, and connections with others are important factors that affect individuals' wellbeing (Becker 1965;Reis et al. 2000).The family distance hypothesis proposes that spouses need to maintain an independent and balanced state of time allocation and communication with others.A short family distance indicates that family members spend more time together, directly squeezing the other party's autonomy and increasing the risk of conflict between spouses.Long family distance leads to estrangement from the family.At the same time, we consider other pressures that accompany family distance.
Existing studies have shown that, due to the emphasis on work and personal responsibility, unemployed individuals are seriously stigmatized in society.Unemployed people are often considered lazy, useless, or unreliable (Brand 2015).In addition to the explicit consequences of earning income, being employed also has significant 'nonmonetary benefits' , including providing a time schedule for a day and defining an individual's status and identity (Jahoda 1981).This means that people with shorter working hours also suffer from stigmatization within their families and the loss of nonmonetary costs.On the one hand, spouses with short working hours may be considered lazy and unsuccessful, and their partner may experience a strong sense of deprivation, while a spouse who works longer hours will be considered hardworking, reliable, or successful, and their partner will develop relative satisfaction.
On the other hand, excessive family distance is also not conducive to improving one's wellbeing.Excessive work by spouses means a decrease in contact with their significant others, and as communication between spouses is an important mediating variable for balancing conflict and marital satisfaction (Carroll et al. 2013), excessive family distance often leads to emotional alienation and the accumulation of conflicts.Meanwhile, if family distance is too great, the spouse may have to assume too much family responsibility, leading to the transfer and imbalance of the family obligation distribution (Bianchi et al. 2000).
Due to word restrictions, this section is just a case demonstration of Computing Grounded Theory and does not use other data to verify the generated theoretical hypotheses. 7Based on the above calculation and analysis results, the family distance theory is summarized; that is, too long or too short of a family distance is not conducive to improving people's wellbeing.We further express this as a series of hypotheses with logical progression.
(1) The family distance of a spouse can affect an individual's wellbeing, but this relationship is nonlinear.(2) A close family distance between spouses can compress one's autonomy time and increase coexistence conflicts.An appropriate family distance will increase one's autonomy and reduce coexistence conflicts.However, long family distances can reduce communication opportunities for family members, leading to emotional alienation and the accumulation of conflicts.
(3) The family distance of spouses affects wellbeing by influencing couple identification.Too close family distance between spouses can lead to a decrease in their 7 We advocate that researchers should use multiple methods to verify theoretical hypotheses after obtaining theoretical inspiration, including cross validation of multiple algorithmic models, cross validation of algorithmic models and data models, cross validation of other data, and cross validation of other theoretical propositions based on logical reasoning that can all be used as methods to validate theories.Researchers should choose appropriate validation methods according to real needs.Meanwhile, we do not suggest that the results of the algorithmic model must fully correspond to the results of the data model.As discussed earlier, when comparing data models and algorithmic models, due to their essential differences in variable quantity, parameter settings, and data relationship pattern assumptions, the relationship between the two models should be complementary rather than competitive.
sense of identification, and coresidence can lead to a sense of deprivation.A moderate or slightly distant family distance can lead to an increase in identification with a spouse and relative satisfaction with coresidence.(4) Family distance affects the distribution of family rights and obligations, and a far family distance can force the partner to take on more family responsibilities.Due to the imbalanced distribution of family obligations, conflicts between the two parties accumulate, thereby reducing wellbeing.

Refined development of theory: exploring multiple patterns of complex relationships
According to the previous section, the curve of the SHAP value of "a spouse's weekly working hours" exhibits a nonlinear pattern.Therefore, is there any common pattern other than this complex relationship between X and Y? We also selected some variables and plotted partial dependence plots of the SHAP values (Fig. 5).It is apparent that we can find many details that regression analysis models cannot discover, and these details are vital for further expanding, supplementing, validating, and clarifying the theory.We obtain five basic patterns of complex relationships with the variation in SHAP values of X.First, there is a "ladder" distribution.The impact of X on Y changes rapidly after a certain turning point and then tends to flatten out, similar to going up a staircase.5c display the partial dependence plots for the variables "self-positioning of class, " "self-expectation of class after 10 years, " and "self-positioning of class before 10 years, " respectively.Fig. 5d, Fig. 5e, and Fig. 5f similarly show the partial dependence plots for "spouse's annual income (CNY), " "personal annual income (CNY), " and "year of birth, " respectively Typical variables include "self-positioning of class", "self-expectation of class after 10 years", and "self-positioning class before 10 years" (Fig. 5a-c).Among them, the key turning point of "self-positioning of class" is 4 (Fig. 5a); that is, if self-positioning is above 4, its impact on wellbeing is positive, and there is no significant difference between classes (the SHAP value is 0.6-0.8).Once the positioning is below 4, it quickly becomes negative (approximately − 0.2), and the impact of lower social classes (1-3) has not changed significantly (maintained at approximately − 0.5).More interestingly, this turning point is different from the expected class turning point for the future (Fig. 5b): the latter has a turning point of 5.This subtle difference means that people who believe they are currently in the middle class of society (= 5) will feel good, but they have higher expectations for the future, while the average marginal contribution of being in the fifth class to wellbeing in the future is only 0.
The second is the "厂" distribution.The impact of X on Y increases sharply in the early stage and tends to flatten out in the later stage, with "personal annual income" and "spouse's annual income" as examples (Fig. 5d-e).This is consistent with the well-being saturation theory: the positive impact of income on well-being tends to have a decreasing average marginal contribution.This has important implications for social governance policies: poverty alleviation programs should allocate limited funds to the most disadvantaged groups.
The third is the "valley" distribution.The impact of X on Y is relatively high at both ends, while in the middle, it is relatively low, forming a valley shape, with a typical manifestation of "year of birth" (Fig. 5f ).The wellbeing of elderly people born before 1955 and young people born after 1995 is significantly greater than that of middle-aged people.In addition, the relationship between age and wellbeing among middle-aged people was not significant, and the SHAP values were almost horizontally distributed.These results echo the groundbreaking literature on age and wellbeing in recent years (Blanchflower & Oswald 2008).
Fourth, there is the "peak" distribution.The impact of X on Y is relatively high in the middle group but gradually decreases on both ends to form a peak, such as the influence of "spouse's weekly working hours" (Fig. 4a) and "face-to-face communication time in the family" (Fig. 4c), which will not be repeated here.
The fifth is the "homogeneity-heterogeneity" effect.The homogeneity effect is manifested as having a consistent impact on the wellbeing of the same group of people, with a small intragroup SHAP variance.The heterogeneity effects manifest as significant differences in the impact on wellbeing among the same group of people, with a large variance in SHAP within the group.Taking "weekly working hours" as an example (Fig. 4b), the SHAP values for working hours ranging from 0 to 40 h are all within − 0.5-0.5, whose distribution is relatively uniform, and the impact of work time on wellbeing is relatively homogeneous.The SHAP values for 70-80 h are distributed between − 1.5-0.1.with a significant effect of heterogeneity on wellbeing.This suggests that people with shorter working hours are generally happier, while those with longer working hours may be happier or less happy, and there may be other important interaction variables that affect wellbeing.

Robustness testing: solving the rashomon effect
Data and algorithms are highly significant in the calculation process.A considerable number of scholars have noted the Rashomon effect of algorithms, namely, the internal heterogeneity caused by different parameter settings and the external heterogeneity caused by different algorithms (Breiman 2001a;Hu et al. 2021).Is there a Rashomon effect in Computing Grounded Theory?To what extent does it exist?This section tests these questions from the following three perspectives.
First, there was heterogeneity in the data.Robust results will not vary significantly with changes in data volume and composition.We use bootstrap self-sampling to randomly select 50%, 60%, 70%, 80%, 90%, and 100% of the original balanced data for calculation and conduct grounded computational theory.
Second, there is heterogeneity in the prediction algorithms.Robust results should be similar under different prediction algorithms.We compare the calculation results of five algorithms: XGBoost, CatBoost, LightGBM, gradient boosting, and random forest.
Third, there is heterogeneity in the algorithm parameters.Different internal parameters of the same algorithm may also lead to different analysis results.We replace the internal parameters of the XGBoost algorithm, including the maximum tree depth (max_depth), regularization coefficient (alpha), learning rate, subsample, and so on.
Under each condition, we obtain a table that includes all features and their means of SHAP absolute values.We calculate the Pearson correlation coefficient for the SHAP results calculated by different conditional models, as shown in Fig. 6.Overall, the training results of these models are highly similar; the correlation coefficients calculated by the pairwise models are basically above 0.95, and the significance of the correlation coefficients is 0.000.The heterogeneity of the data and the internal heterogeneity of the algorithm parameters are basically nonexistent.There is a certain degree of heterogeneity in the prediction algorithms, but the minimum also reaches above 0.88.We also calculate the Spearman correlation coefficient based on rank, and the analysis results are highly similar to the Pearson coefficient, so we do not report them here.In brief, in the case of wellbeing, Computing Grounded Theory has a considerable degree of robustness.

Recommended technical standards for grounded computational theory
No unified standard for machine learning training methods has been applied in the field of social sciences.Therefore, we selected 60 highly cited papers in the social sciences from the core collection of the Web of Science with the search keyword "machine learning" and summarized information such as the number of variables, sample size, model selection, and model evaluation indicators commonly used in training models, providing empirical reference standards for algorithmic model training. 81) Number of samples.According to the statistical results of the literature, the median sample size was 1,888, and the median sample size after 2015 was 11,196.After ensuring sample availability and representativeness, we suggest that the sample size for exploratory analysis should be greater than 2000, and the sample size for exploratory and validation analysis should be even larger.
(2) Sample balance.The sample size should be adjusted according to the number of categories and difficulty level of the variables to be predicted.It is particularly worth noting that the sample size for rare categories should not be too small.For imbalanced samples, resampling should be used to balance the various sample sizes (Chen et al. 2022).
(3) The number of features, that is, the number of X variables used for training.
According to the literature, the largest number reached 1,821, and the median number of all papers was 22.5.Many X variables will bring better training results and be conducive to discovering new potential theories.However, it is also important to consider that certain models may be sensitive to data noise.(4) Algorithmic model selection.Among the 60 papers, the most commonly used algorithm was random forest (29%), followed by support vector machine (26%).Neural network and gradient enhancement algorithms account for approximately 17% and 15%, respectively.Most papers have adopted more than one algorithm and compared the performance results of the models.Therefore, we suggest comparing the predictive performance and results of multiple algorithms, selecting the optimal model as much as possible, and conducting robustness tests.(5) Model prediction performance.The median accuracy of the model in the existing literature is 0.79. 9The effectiveness of Computing Grounded Theory is based on the accuracy of model predictions.Given that most of the predicted variables in existing papers are binary variables, we suggest that the accuracy of binary variables should be greater than 0.8.This criterion for the accuracy of continuous variables can be appropriately reduced.

Multiple values of computing grounded theory
The development of Computing Grounded Theory to traditional quantitative research is multifaceted.At the data level, the indicators included in the model are no longer a limited group of variables; rather, as many indicators are included as possible.At the target level, the statistical significance of model coefficients is no longer emphasized, but the accuracy of social prediction and the interpretability of mechanisms are rebalanced.
From the perspective of observation, this approach is no longer limited to the number and direction of regression coefficients but rather involves careful exploration of the nonlinear relationships and population heterogeneity effects between variables.These innovations can bring various values to grounded computational theory.

The value of theoretical generation: discovering potential patterns
Compared to traditional data models, algorithmic machine learning methods can overcome the limitations of model form and variable selection and consider various interaction relationships among variables.The breaking of the bottleneck in the number of independent variables and the limitation of relationships allows us to gain a more robust ability to search, think, and test explanatory variables.This means that as long as the data itself are abundant, Computing Grounded Theory can guide researchers to form new theoretical hypotheses by discovering new explanatory variables (Chen et al. 2020(Chen et al. , 2021)).For a given dataset, by performing the method of Computing Grounded Theory only once, we can filter and compare hundreds of indicators across the entire survey dataset.

Value of theoretical development: capturing complex relationships
Traditional econometric methods use data fitting models, which can easily lead to the loss of, or even errors in, key information (Varian 2014).Grounded computational theory uses hyperparameters to fit data.As long as the model simulates real social situations as much as possible, it will fully capture the complex relationships between variables, liberate the linear shackles of traditional econometric models, and validate or develop theories.The previous case clearly demonstrates its ability to reveal and explain complex relationships and reminds us that the pairwise relationship of data in the real world is far from as neat and uniform as we expected: the SHAP curve has almost not obeyed any distribution of a straight line.

The value of the discipline paradigm: the second sociological imagination
Hofman and Watts et al. published an article in the 2021 issue of Nature calling for the integration of interpretation and prediction in the computational social sciences (Hofman et al. 2021).They noted that research activities that integrate interpretive and predictive thinking are valuable, but existing research is scarce, and this field should receive more attention than it has so far.This method is a new attempt to integrate the explanatory and predictive aspects of social science.For quantitative research paradigms, mastering Computing Grounded Theory is equivalent to obtaining a supplement beyond the sociological imagination proposed by Mills.Mills' sociological imagination is the enhancement of an analytical perspective based on personal experience (Mills 2017), while Computing Grounded Theory provides a data-based thinking ability to directly assist in the production of theory through algorithmic models.The second type of sociological imagination contains the immense power that drives new theoretical discoveries and unleashes potential.

The value of knowledge systems: autonomous knowledge production
Grounded computational theory naturally has stronger power in the production of systematic knowledge: there are many new theoretical hypotheses that can be extensively inspired by data and more subtle mechanisms and relationship features that can be simultaneously discovered through detailed contribution analysis of predictive power for theoretical development and clarification.A real autonomous knowledge system requires a tool that is capable of discovering complex relationships and extracting theories from large-scale, spatiotemporal, and high-throughput data.Computing Grounded Theory is undoubtedly one of the most important components of such tools.

The value of social governance: finding intervention factors
Sociology is a practical science, and the public and governance entities are often not satisfied with concept extraction, process interpretation, or statistical judgment.This means that the disciplinary mission of quantitative sociology cannot be confined to verifying theoretical hypotheses; but it must also master the ability to identify key intervention factors for social phenomena to provide advice and suggestions for those who serve the country.Computing Grounded Theory is a problem-oriented approach to social prediction, and by simulating existing social phenomena, it has important practical value in identifying key intervention variables for social governance.

Conclusion
As an important research paradigm in the field of sociology, quantitative research is deeply rooted in the tradition of positivist methodology and tends toward single-path dependency on hypothesis testing.The overemphasis on theoretical validation in quantitative research may overlook the enormous theoretical energy inherent in the data itself.Based on this, this article proposes a method for theoretical production based on quantitative data: Computing Grounded Theory.With the predictive power of machine learning and the interpretability of attribution algorithms, Computing Grounded Theory can open up a gate for developing quantitative research in terms of exploring potential relationship patterns and capturing nonlinear relationships and for opening up the path from empirical observation to theoretical production.With respect to the development history of quantitative methods, more than forty years ago, there was an academic trend in the social sciences of exploring the relationships among variables based on data and models; however, no mature research paradigm had yet been developed.The reason for this is that the inclusion of many variables makes it difficult to satisfy assumptions and leads to problems such as multicollinearity.Second, the method of selecting variables by deleting or adding a single indicator is only a partial rather than a globally optimal solution, and changes in control variables can cause significant disturbance to the results.Third, the presupposed function mode has difficulty exhausting the complex relationships and action modes among variables.Forty years later, we once again call for quantitative research to fill the gaps in theory production, learn from historical lessons, and address the enormous theoretical energy contained in the data.The advantages of grounded computational theory include the following: first, the algorithmic model can break the assumptions and relationship patterns made by the model, fully incorporate a large number of variables and consider the complex relationship patterns among variables; second, interpretable machine learning can utilize the algorithmic power to obtain the global optimal solution while fully considering various permutations and combinations of variables; third, the importance ranking of variables based on predictive power is closer to the category of causal relationships in analytical logic than simple variable correlations and; fourth, mining and visualizing various nonlinear relationship patterns among variables provide more solid and detailed information for guiding theoretical intuition.
While celebrating the capability of algorithms and data in theory generation, it is worth noting that this article does not negate "traditional" quantitative methods and their value.Each method has its own premises, assumptions, and limitations.They are all important components of quantitative sociological methods.We emphasize that Computing Grounded Theory is not a rejection of theory but rather a departure from the limitations of existing theories and common sense, creating opportunities for proposing new hypotheses.Computing Grounded Theory does not exclude the validation of theories but emphasizes the generation of theories from data as a scientific step before quantitative scholars test theories.
We are fully aware that new analytical methods often propose new research questions.The challenge of Computing Grounded Theory is as significant as the knowledge production it can bring.These challenges include the following: first, the limitations of the data.Variables can never be exhausted.Although Computing Grounded Theory is trying to broaden the data dimension for analysis, it is not far from "dancing with the shackles of data availability".Second, there are limitations to social prediction.There has always been skepticism about the predictability of complex social phenomena (Taleb 2010).Due to insufficient data, model limitations, and the inherent unpredictability of complex social systems, grounded computational theory is not applicable to all research scenarios.Third, there is heterogeneity in grounded computational theory.The position of knowledge production for researchers shifts from the front-end to the back-end, with data and models being pushed to a significant position and potentially leading to potential biases.Fourth, correlation is not

Fig. 2
Fig. 2 Overall path of Computing Grounded Theory based on machine learning.Legend: Fig. 2 illustrates the six basic steps of Computing Grounded Theory.The first two steps are preparatory steps, and the third and fourth steps represent core technical steps involving the combination of machine learning and interpretable machine learning.The final two steps focus on theory generation

Fig. 3
Fig.3Top 20 factors with the largest average marginal contribution (SHAP value) to wellbeing prediction.Legend: In 3a, the average absolute value of the SHAP value for each variable (X) across all cases is depicted, representing the average marginal contribution of each variable.Figure3bshows the details of the influence of different predictive variables in the form of scatter plots.Each point in Fig.3brepresents a real sample.For each row, the color represents the magnitude of feature X of the variable in that row.The larger the X is, the darker the color of the point.The horizontal axis represents the size of the SHAP value.The more points with the same SHAP value, the larger the cross-sectional area of the honeycomb and the thicker it appears.Taking the sense of fairness as an example, the scatter plot shows that cases with a greater sense of fairness (black dots) often concentrate on the right side of the horizontal axis; that is, a positive SHAP reflects an increase in happiness, while cases with a lower sense of fairness (gray dots) often concentrate on the left side of the horizontal axis, where the SHAP is negative and happiness is thus reduced.This indicates that a sense of fairness has a typical positive impact on happiness

Fig. 4
Fig. 4 Partial dependence plots of the new theory of family distance.Legend: The gray dots represent samples; the horizontal axis represents the true value of the relevant feature X of the sample; the vertical axis represents the average marginal contribution of X corresponding to the sample, which is the SHAP value.The black line represents the line of the mean SHAP value of X at each value, and the change in the line reflects the change in the relationship between the two variables.Fig. 4a, Fig. 4b, and Fig. 4c respectively reflect the partial dependence plots for the variables "spouse's weekly working hours", "weekly working hours" and "face-to-face communication time in the family"

Fig. 5
Fig. 5 Partial dependence plots of the SHAP values of related variables.Legend: The gray dots represent samples; the horizontal axis represents the true value of the relevant feature X of the sample; the vertical axis represents the average marginal contribution of X corresponding to the sample, which is the SHAP value.The black line represents the line of the mean SHAP value of X at each value, and the change in the line reflects the change in the relationship between the two variables.Fig. 5a, Fig. 5b, and Fig. 5c display the partial dependence plots for the variables "self-positioning of class, " "self-expectation of class after 10 years, " and "self-positioning of class before 10 years, " respectively.Fig. 5d, Fig. 5e, and Fig. 5f similarly show the partial dependence plots for "spouse's annual income (CNY), " "personal annual income (CNY), " and "year of birth, " respectively

Fig. 6
Fig. 6 Results of the robustness test of the computing grounded theory.Legend: The X and Y axes represent 21 different scenarios, including different data, models, and parameters, for which SHAP values were calculated.The value in each small square represents the Pearson correlation coefficient calculated for each pairwise combination of these scenarios, with color indicating the magnitude of the coefficient.Lighter colors indicate greater similarity between the SHAP results calculated by the models for these two scenarios Two logics of Wallace's the Wheel of Science.Legend: This figure is adapted from Wallace, Walter L. 1971.The logic of science in sociology, p. 18, with simplifications.The Wheel of Science illustrates that sociological research is an endless, spiral process that includes theory construction and testing ). 2 Moreover, relevant empirical research has combined machine learning and interpretable AI methods, such as identifying potential factors that lead to anxiety (Nemesure et al. 2021) and exploring potential variables that improve the survival rate of patients with breast cancer (Moncada-Torres et al. 2021).Chinese scholars, such as Jiade Luo, have used relevant methods to revise and clarify the theoretical model of Chinese network circles (Luo et al. 2021), while Zhou and coauthors (