Improving Cognitive Scaffolding in Computer Science Education

The current pedagogy currently places a strong emphasis on unguided problem solving and independent experimentation to help students improve their programming skills. Discouraging intellectual hand-holding in favour of having students learn from their own mistakes can be interpreted as a "No pain, no gain" theory where depth of learning stands in proportion to student effort. Taken to an extreme, it implies that classroom instruction is redundant, and therefore research into computer science (CS) education that assumes this dominant pedagogy tends to focus on developing expensive interactive hypermedia such as micro-worlds, algorithm visualisations and visual debuggers.

Reviews of the pedagogical efficacy of these technologies indicate however that the best predictor of learning is not the number of interactive affordances, but how cognitively well-structured the contents are. In other words, what matters appears to be that the student's attention is directed, or scaffolded, by visual and verbal means, to ensure the formation of an accurate mental model. If explanations were maximally clear and specific, then maybe time-costly iterations of trial and error would not be necessary.

Hence, the literature identifies two overarching frameworks for understanding learning in CS: one is a problem-solving first approach; the other advocates for the development of appropriate underlying mental models. They are both so high-level it is difficult to settle experimentally which one is superior and when. Instead, this project aims to home in on specific scaffolding techniques that can accelerate the novice's process of understanding, designing and implementing algorithms. It investigates the hypothesis that novices benefit from low-level cognitive scaffolding specifically in the abstraction of patterns and construction of schemas.

In more practical terms, the project will build up a catalogue of examples of structural patterns among algorithms, label them, represent them graphically and in pseudocode, and annotate them with common errors and ambiguities. The latter would require qualitative methods of interviewing and testing novices to reveal which aspects are most prone to confusion. It will also require designing tests to measure fluency with simulating and implementing the algorithm.
It will then use these to design computerised algorithm tutorials varied on their degree of scaffolding, structure or other cognitive parameters. A concrete example would be collecting post-order traversal implementations for different ways of representing a tree data structure and, in one tutorial variant but not the other, explain the underlying invariance (e.g. parent node stored inside node or not). This could be used to test the hypothesis that explicit many-to-one mapping between concept and implementations makes the understanding more transferable, and less over-trained to a particular code surface appearance.
To assess efficacy, the project will recruit undergraduates and maybe also secondary school students. Using a matched between-subjects design, it will administer the tutorials, then measure short-term and long-term effects. Because it aims to maximise a specific outcome, it could be described as a form of the A/B testing schema currently popular in web analytics. It will additionally collect questionnaire data on the effects of scaffolding on student confidence and self-perception, to inform discussions on whether it could reduce the high drop-out rates in CS college courses.

With an increasing demand for programming skills combined with widely reported low success rates in CS education, this project could go far in innovating and optimising ways of structuring content in the teaching of algorithms, which would be much cheaper and potentially more effective than the interactive hypermedia focused on until now.