Introduction to Adaptive Learning
Using machine learning and data science to personalize education
Data Science (DS) and Machine Learning (ML) are often leveraged to build personalised products. The application of personalization to education is what encompasses the field of Adaptive Learning. It’s a relatively less commercialised application of DS/ML that has attracted top researchers in the field including Tom Mitchell (yes, THE Tom Mitchel) to run companies dedicated to solving it.
In this post we will dive into what adaptive learning is, foundational concepts, design of adaptive learning systems, and DS/ML techniques that are commonly used in the field.
What is Adaptive Learning?
Adaptive learning aims to personalise course instruction for individual learners. In an era where the teacher to student ratio is ever dwindling, there is a huge amount of interest in creating learning systems that can support both teachers and students. Adaptive learning aims to support students by providing real-time feedback and adapting to students’ learning nuances. In the case of teachers, adaptive learning can help identify students that are at risk of falling back, and provide insights on how students’ learn concepts to improve the course over time.
Intelligent Tutoring Systems
A solution that leverages adaptive learning is commonly called an Intelligent Tutoring System (ITS) in the literature. A typical ITS is depicted below.
From the numbered and highlighted segments we see that an ITS adapts to students’ needs in three different ways.
Design Loop Adaptivity
ITS can collect data from one cohort of students that interacted with a course and provide it to the instructors, who can then better design the course for the next cohort. This is called design loop adaptivity, where the course is adapted based on data to entire cohorts at a time.
Task Loop Adaptivity
ITS can collect an individual learner’s performance, as they interact with course content. Then based on a ‘domain model’ (model of information that connects course concepts), a ‘learner model’ (a model tracking what set of concepts a learner has learned thus far and with what confidence), and a ‘pedagogical model’ (model of what should be taught next to a learner who has learned a certain set of concepts with certain confidence) the ITS can adapt the next set of course instructions, called task loop adaptivity. Thus, task loop works to personalise course content for individual learners.
Step Loop Adaptivity
ITS can also adapt to learners’ individual actions within a task. For instance, a student working on a task can receive real-time feedback on if they are performing intermediate steps in the task correctly or not, ask for hints on the next step, understand repercussions of intermediate steps etc. This is called ‘step-loop’ adaptivity. The step loop works to adapt within a task, to the actions by an individual learner.
What aspect of learning to adapt to?
So now that we understand the levels at which we can build adaptivity into a system, what all aspects of a students’ learning can we adapt to? Should we limit ourselves to just student knowledge, given that intuition would tell us students can have different learning styles and associate different emotions with learning content? Following are some of the possible elements we can adapt to, the DS/ML techniques that can be used, and the research-based conclusion surrounding them thus far.
Prior Knowledge and Knowledge Growth
Learners have varying degrees of prior knowledge and experience in a target domain, so one-lesson-fits all approach doesn’t work. As such an ITS needs to be able to assess what students know and adapt instruction accordingly. Based on current research, this can be effectively done at all three level of adaptivity
Design Loop Adaptations: The course interaction and performance data are analysed to identify insights such as missing prerequisites students need to be trained on, incorrect assumptions students might be making based on incomplete instruction etc. The field of educational data mining deals with such multidimensional analysis at different scales.
Task Loop Adaptations: At this level, we attempt to predict the next best task a learner should work on based on models we create about his knowledge thus far. A popular approach uses Cognitive Mastery for Task Selection and Bayesian Knowledge Tracing models to determine what is the ideal next task to maximise learning while minimising redundant effort.
Step Loop Adaptations: At this level, we attempt to provide real-time feedback to learners as they are working through a task e.g. correctness on intermediate steps. Research indicates this helps reduce learner uncertainty and thus increase learner efficiency. A simple and popular approach to enable this is rule-based cognitive modelling.
Strategies and Error
Learners may use different strategies to achieve the same goal while working through a task (like using different methods to solve a maths problem) and this may result in learners being confronted with different errors. (like calculation mistake, unit conversion error, incorrect formula etc) Research indicates ITS can adapt to such learner preferences at the design and step loop levels. Currently, there is no conclusive research of any task loop adaptation to this goal that improves learning.
Design Loop Adaptations: At this level, we attempt to ensure students are able to use different strategies, but at the same time require enough intermediary steps, also called ‘scaffolding’, so that we can track what strategies they are using without explicitly asking them about it. Based on the interaction and performance data, instructors may decide to change scaffolding around tasks to allow for more strategies and to minimise errors. Knowledge component modelling is a popular approach to account for such design loop adaptations.
Step Loop Adaptations: At this level, we attempt to explain errors, or repercussions on erroneous steps, and provide next-step hints based on student strategies that we have deduced thus far. Dynamic Bayesian Networks and Reinforcement Learning are common approaches to build step loop adaptations.
Affect and Motivation
Learners can experience different emotions such as boredom, confusion, frustration, engagement/flow, also known as ‘affective states’, which can influence how they perceive course content and perform in general. As such an ITS must be able to detect affective states, and promote those that are conducive to learning. Research indicates ITS can adapt to affective states at all three levels of adaptivity.
Design Loop Adaptations: At this level, interaction and performance data is analysed to understand affective states that are conducive to learning. ITS is designed to promote learning-conducive affective states. For example, educational data mining is often used to identify learners and behaviours that are related to ‘gaming the system’, where learner appears disengaged (identified via facial cue detected using computer vision), and maybe taking advantage of real-time feedback ITS provides e.g. correctness feedback on intermediate steps or next-step hints. Instructors can then design mechanisms to discourage this.
Task Loop Adaptations: At this level, we aim to select the next task that will not only optimize learning performance but also generate affective states conducive to learning. For example, A/B testing indicated tailoring task context around student interests such as sports/entertainment/art etc lead to better more student engagement and overall better performance.
Step Loop Adaptations: At this level, we aim to provide support in-task that promotes positive affect states. This can be in the form of empathic dialogue when providing hints, alerts to bring back distracted students, or motivational messages. Research and A/B testing indicates this causes students to perceive ITS in a better light, and achieve better learning outcomes.
Self-Regulated Learning (SRL)
Learners take several self-motivated actions to grasp concepts such as setting goals, contrasting different strategies and assessing their own performance to know what their strengths and weaknesses are. All these actions fall under SRL and it accounts for substantial variability in learning outcomes. Research indicates SRL can be adapted at design loop and step loop levels. The research for task loop adaptation is inconclusive.
Design Loop Adaptations: At this level we leverage educational data mining and user surveys to understand SRL behaviours that lead to optimal learning. Instructors then design mechanisms in the ITS to promote successful SRL behaviours. e.g. Research and A/B testing indicate versions of ITS that ask students to explain their answers lead to better learning outcomes.
Step-Loop Adaptations: At this level we leverage the domain and learner models to estimate student proficiency in-task, and provide adaptive messaging to encourage SRL activities. For example, based on student interaction with tasks — time taken, attempts, hints used thus far etc, students that are struggling can be identified. Such students can be provided adaptive messages, to reflect and explain what they have done thus far before proceeding. Research indicates such step-loop adaptivity leads to better learning outcomes and promotes ‘preparation for future learning’.
Learning Styles
Learners have various self-reported preferences for how they learn. Now, intuition would say that adapting content to match students’ learning style would improve learning outcomes. However research thus far provided no evidence for the same at any level of adaptivity.
As such, summarising research we can create an overview of what we can adapt to vs the level at which we adapt (design/task/step), referred to as the adaptivity grid in literature. (depicted below)
Conclusion
In this post we covered what adaptive learning is, what an ITS system looks like, the different level at which we can adapt an ITS, and the different elements of learning to which we can adapt an ITS.
For those looking to delve deeper into this relatively new field, apart from the domain specific techniques (Cognitive Modelling, Knowledge Component Modelling, ACT-R) the key DS/ML knowledge required includes Reinforcement Learning, Bayesian Modelling, and A/B Testing.
