Learning Analytics to Support Teachers’ Assessment of Problem Solving: A Novel Application for Machine Learning and Graph Algorithms

In contrast to well-structured problems which have pre-defined, correct answers, complex real-world problems are often ill-structured problems (ISPs). The open-ended nature of ISPs creates considerable barriers to assess and guide students in forming better solutions, which results in low adoption levels of inquiry-based learning. Students can structure and represent their knowledge for an ISP in the form of knowledge maps or causal maps, which articulate relevant concepts and their causal relations (i.e., antecedents and consequents). Assessing such maps can involve a referent-free evaluation (e.g., to encourage the creation of maps with high density of concepts) or a comparison to an expert map used as reference. This chapter starts with a review of theories and tools to compare a student’s map to the expert map. Previous approaches often compared individual connections (e.g., scoring the number of connections that a student has/misses in contrast with the expert) or general map metrics (e.g., one map is denser than the other). In contrast, the problem of comparing two maps has been studied in network theory and graph theory for several decades, yielding categories of algorithms that are currently underutilized in educational research. This chapter reviews three categories of algorithms (i.e., graph kernel, graph editing distance, graph embedding) in light of their application to assessment and student success. We discuss an implementation of these algorithms through a new set of digital tools, designed to support a community of practice in problem-based instruction.