Modeling and predicting students' engagement behaviors using mixture Markov models

Rabia Maqsood, Paolo Ceravolo, Cristóbal Romero, Sebastián Ventura

Introduction

In digital learning environments, student engagement is far from static. It is a dynamic, evolving process where a student might transition from deep focus to frustration or disengagement within a single session. Understanding this "drift" is crucial for developing adaptive educational systems that can provide timely interventions. This paper introduces a novel approach to modeling these transitions using Mixture Markov Models (MMM) enhanced by a specialized initialization technique called K-EM.

The Challenge of Modeling Engagement

Traditional models often struggle to capture the diversity of behaviors across a student population. While Mixture Markov Models are well-suited for grouping similar behavioral sequences into clusters, they rely on the Expectation-Maximization (EM) algorithm for training. EM is notoriously sensitive to its starting parameters; random initialization often leads the model to settle into local optima rather than finding the most accurate global solution. This sensitivity is particularly problematic in educational datasets where interaction sequences may be short or sparse.

Methodology: The K-EM Approach

The researchers developed a multi-stage pipeline to more accurately model and predict engagement.

Behavioral Mapping

The first step involves translating raw interaction data—such as answer correctness, self-reported confidence, and feedback-seeking behavior—into discrete behavioral patterns. These patterns are categorized into "Engaged" and "Disengaged" states.

Mapping of student actions to behavioral engagement patterns.

As shown in the mapping table, behaviors like "High Knowledge" (HK) and "Fill-knowledge Gap" (FG) represent engaged states, while "Knowledge Gap" (KG) and "Not Interested" (NI) signal disengagement. By incorporating confidence and feedback-seeking, the model distinguishes between a student who is struggling but persistent (FG) and one who is simply guessing (KG).

The K-EM Algorithm

The core contribution of the paper is the K-EM initialization method. Instead of starting the EM algorithm with random values, the authors perform a preliminary K-means clustering on the proportional counts of a student's behaviors. The results of this computationally "cheaper" clustering are used to set the initial transition matrices and weights for the more complex Mixture Markov Model. This "warm-up" phase ensures the EM algorithm starts in a more promising region of the parameter space.

Experimental Setup and Datasets

The study validated the K-EM approach using two distinct real-world datasets:

  1. Dataset 1 (Short Sequences): 94 freshmen in Pakistan performing code-tracing tasks, with an average of 5 problems per session.
  2. Dataset 2 (Long Sequences): 210 undergraduate students in Italy working on flow diagram multiple-choice questions, with an average of 17 problems per session.

Key Results and Insights

The K-EM method demonstrated superior performance across several metrics:

  • Higher Accuracy: On the sparser Dataset 1, K-EM achieved a Macro Accuracy of 62.26%, significantly outperforming the non-mixture baseline (53.48%) and standard random-start EM (56.61%).
  • Robustness: The improvement was most pronounced in the dataset with shorter interaction traces, suggesting that K-EM is particularly valuable for "sparse" data where patterns are harder to extract.
  • Actionable Visualizations: By representing clusters as Markov Chain diagrams, the authors provide a way for educators to visualize the "flow" of a classroom. For example, teachers can identify clusters of students who successfully transition from "Less Knowledge" to "High Knowledge" after receiving feedback, versus those who remain stuck in disengaged states.

Limitations and Future Directions

While the results are promising, the authors note that the current model uses a first-order Markov assumption, meaning a student's next action depends only on their current state. Future work may explore higher-order Markov chains to capture more context, though this increases computational complexity. Additionally, the researchers aim to utilize Hidden Markov Models (HMMs) to allow for "intra-session" transitions, where a single student can move between different engagement profiles as they progress through a task.

Summary updated

BibTeX
@article{maqsood2022modeling,
  author = {Rabia Maqsood and Paolo Ceravolo and Cristóbal Romero and Sebastián Ventura},
  title = {{Modeling and predicting students' engagement behaviors using mixture Markov models}},
  journal = {Knowledge and Information Systems},
  year = {2022},
  doi = {10.1007/s10115-022-01674-9},
}
Figure