Learning path optimization with incomplete learning object metadata

Abstract

One of the fundamental concerns of instructional design is pedagogical sequencing which is a practice of organizing course materials according to the underlying knowledge structure and concept dependency. In the conventional settings, like the secondary schools or tertiary institution, instructors are required to interpret learning materials by their own domain knowledge. But in many online learning systems, analyzing and interpreting learning materials are very challenging due to the lack of instructional contexts and pedagogical attributes of the learning units.

The learning objects and learning object metadata (LOM) are learning technologies to formalize the concept of learning unit and standardizing the specification of learning object annotation framework. The learning object is aimed to provide a solution for reuse and sharing of learning materials, and to provide infrastructure for pedagogical design. The LOM has been widely adopted in various learning systems, methodologies and system frameworks proposed to solve instructional design problem based on the pedagogical information as provided in the LOM. However, an empirical study showed that most real-life learning objects do not provide necessary pedagogical information. Thus, it is not clear how the issue of incomplete metadata and hence incomplete pedagogical information will affect those LOM based methods.

A new approach to reconstruct the underlying knowledge structure based on information extracted from LOM and data mining techniques is proposed. The main idea of the approach is to reconstruct knowledge structure by the context of learning materials. Intrinsically, the vector space model and the k-means clustering algorithm are applied to reconstruct the knowledge graph based on keyword extraction techniques, and concept dependency relations are extracted from the obtained knowledge graph. Then, the genetic algorithm is applied to optimize for a learning path that satisfies most of the obtained concept dependencies. Furthermore, the performance of applying different semantic interpreters and rule extraction methodology are carefully tested and compared. Experimental results revealed that learning paths generated by the proposed approach are very similar to learning paths designed by human instructors.