2.1 Human-Machine Collaboration
Contemporary human-machine collaboration refers to the synergistic partnership between humans and intelligent machines. These machines can take various forms, including automated systems, autonomous agents, robots, algorithms, or artificial intelligence (AI) entities [15, 37]. This collaborative approach results in enhanced performance by leveraging the strengths of both intelligent machines and human intelligence, addressing their respective limitations [2, 34, 38]. Intelligent machines also excel in processing vast amounts of information and generating rational outcomes without succumbing to cognitive biases (e.g., availability bias, representativeness bias, and anchoring effect) or being swayed by internal and external factors (e.g., ability, cognitive style, emotions, workload, fatigue, and time pressure) [32]. Conversely, humans possess unique advantages in employing intuition and experience to discern critical factors, adapt to novel conditions, and rapidly learn and apply reasoning to navigate high uncertainty or tackle new, complex, and rare challenges.
In light of these benefits, human-machine collaboration has experienced a growing application across diverse domains, including engineering [29], healthcare [21, 31], and business [11, 15]. For example, a study by Wilson and Daugherty [33] analyzed 1500 companies spanning 12 industries and found that the most substantial performance enhancements occurred when humans collaborated with machines. Moreover, research indicates that human-machine collaboration surpasses the efficacy of operations involving only humans or machines separately. For instance, a study by Xiong et al. [36] explored the performance of human-only, machine-only, and human-machine joint teams in a sequential risky decision-making task. The findings revealed that the human-machine joint teams outperformed both human-only and machine-only teams. In the human-machine joint team, the machine as a partner entailed human decision-makers to cede power and coordinate, and their pumping decisions became more conservative and fluctuating.
The trajectory of human-machine collaboration also extends to medical disciplines. A noteworthy example of successful human-machine collaboration in healthcare is evident in cancer detection through the analysis of lymph node cell images [31]. The study demonstrated that combining predictions from a deep learning system with diagnoses from a human pathologist achieved an area under the receiver operating curve (AUC) of 0.995, surpassing the AUC of the deep learning system alone (0.925) and that of the pathologist alone (0.966). This integration resulted in a remarkable reduction in error rates, amounting to at least 85%. Beyond cancer detection, collaborative frameworks have been instrumental in areas such as personalized medicine, where the integration of machine-generated insights with clinical expertise allows for tailored treatment plans based on individual patient characteristics (e.g., Khan et al. [12]).
2.2 Explainable AI
The human-machine collaboration can be enhanced by enabling machines to explain their reasoning in a way that is understandable and trustable to humans [6]. This can be achieved through the integration of XAI [22]. XAI, a sub-field of AI, provides human-interpretable explanations regarding the rationale, strengths, weaknesses, and anticipated behavior of AI systems [9, 23]. In recent years, the significance of XAI has increased due to the widespread applications of advanced AI techniques such as deep learning models. Despite their remarkable accuracy in predictions and classifications, these models are often characterized as "black box" models [23, 27]. This label stems from the reliance of machine learning models on mathematical constructs, featuring an extensive array of abstract, numerical parameters, often numbering in the millions or even billions. These parameters are learned from training data, presenting a challenge in offering profound insights into the intricate dependencies, causal relationships, and internal structures of the models [3, 18]. The opaqueness inherent in these black box models introduces the potential for misleading users [20], raising substantial concerns, particularly in sensitive domains such as healthcare and other applications that involve human life, rights, finances, and privacy [3].
To enhance the interpretability of AI outputs, researchers have proposed various XAI methods. According to the latest comprehensive review conducted by Minh et al. [16], XAI methods fall into three main categories: pre-modeling explainability, interpretable models, and post-modeling explainability. The pre-modeling explainability method involves a set of data processing approaches applied to gain insights into datasets used for training ML models. This includes data analysis, summarization, and transformation. On the other hand, interpretable models refer to those that can be understood by humans through examination of the model summary or parameters, such as linear models, decision trees, k-nearest neighbors, and rule-based models. Lastly, the post-modeling explainability method aims to enhance the interpretability of existing black-box ML models by employing various techniques.
Given its widespread application, Minh et al. [16] categorized post-modeling explainability techniques into four main types. First, textual justification generates explanatory text in the form of phrases or sentences. Second, visualization provides clarity through visual images, utilizing techniques like layer-wise relevance propagation (LRP) and local interpretable model-agnostic explanation (LIME). Third, simplification creates a new and simpler system from complex ML models, employing techniques such as local explanation and example generation. Fourth, feature relevance quantifies the importance of input variables, incorporating techniques like SHapley Additive exPlanations (SHAP). According to Minh et al.'s [16] summary, visualization, simplification, and feature relevance emerge as the three commonly used XAI methods, emphasizing their role in rendering AI systems more transparent and understandable.
2.3 Predicting High-School Dropout
The issue of high school dropouts has long been a focal point in education. For example, research conducted in Wisconsin revealed that approximately 3,000 students discontinue their education before reaching the 12th grade, with around 1,500 of these dropouts occurring during the 9th and 10th grades (Knowles, 2015). In response to this concerning trend, efforts by researchers, policymakers, and school administrators have been directed toward the development of early warning systems powered by ML models. These systems aim to identify students at risk of dropping out of high school and uncover actionable predictors that can guide future interventions and policy adjustments (Allensworth et al., 2018; Bowers, 2021).
To date, a plethora of studies have harnessed ML models to forecast high school dropout based on various background and demographic characteristics exhibited by students, including low grades, aggressive behavior, student poverty, and high absenteeism. For instance, Sara et al. (2015) employed the Random Forest algorithm to predict the dropout status of Danish high school students, utilizing demographic and school-related variables such as gender, school and class size, and teacher-pupil ratio. Chung and Lee (2019) similarly utilized RF to anticipate the dropout status of Korean high school students. In contrast, Sansone (2019) delved into the dropout phenomena among American students, employing Support Vector Machine, Boosted Regression, and Post-LASSO algorithms. Interestingly, this study discovered that GPA, rather than demographic variables, emerged as the most accurate predictor.
While ML models have been extensively employed in dropout prediction, only a limited number of studies have integrated XAI to comprehend high school or college dropout [14, 15, 19]. For example, Krüger et al. [14] investigated dropout factors within the Brazilian technical school system using XAI methods, specifically SHAP and LIME. The findings highlighted the significance of the year of elementary school completion, the family's minimum wages, and the mother's education and work characteristics as important predictors of dropout. Additionally, Nagy and Molontay [18] also employed XAI techniques, SHAP and LIME, revealing that a higher GPA in high school or higher marks in the mathematics section of the matura exam could significantly reduce the likelihood of college dropout.
A significant drawback in prior research exploring high school dropouts through ML models or XAI lies in the substantial reliance on immutable predictors rather than actionable predictors. Immutable predictors encompass variables over which students, teachers, administrators, and family or community members possess limited or no control—examples include gender, ethnicity, and socioeconomic status. On the other hand, actionable predictors, also known as malleable predictors, denote variables that are recent or real-time, adaptable, and amenable to intervention. These predictors can be utilized to implement tailored interventions or modify the current education system. Examples of actionable predictors include orientation to the future and academic habits of mind, such as self-regulation, self-efficacy, and time management (Ben-Avie & Darrow, 2018; Bowers, 2021).
2.4 Current Study
While prior studies have made significant contributions to the domains of human-machine collaboration, explainable AI, and high school dropout prediction, there remain notable gaps that warrant further exploration. Firstly, despite the burgeoning use of human-machine collaboration in fields such as engineering, business, and healthcare, its application within the context of education remains underexplored. However, the potential benefits of incorporating human-machine collaboration in education are substantial. This approach has the capacity to enhance efficiency and accuracy, allowing educators to dedicate more time to personalized teaching methods. Furthermore, the integration of AI can yield results that are more user-friendly, ultimately assisting teachers in improving student engagement and achievement. Therefore, this study seeks to exemplify the implementation of human-machine collaboration in education, using high school dropout prediction as a case study. Secondly, previous research on predicting high school dropouts through ML or XAI techniques has primarily concentrated on achieving higher prediction accuracy based on immutable predictors. However, these immutable predictors offer limited guidance for conducting interventions or modifying the current education system. Consequently, this study aims to address these dual gaps by identifying actionable factors for predicting high school dropouts through the incorporation of the human-machine collaboration paradigm.