Description of work
T3.1 – Learners’ future behavior prediction (GIO, UoM, AAU) [M4-M16]
The task is dedicated to:
- Building datasets for the machine learning models
- Learn machine learning models for different purposes and optimize their parameters
- Make predictions accessible
Using appropriate metrics and features from past user behavior, models can be constructed which predict the future learning performance in unseen tasks or the performance of other students on the same task. For example, having observed several students completing an entire virtual lab, one can make prediction if a student will successfully complete a virtual lab early on. In case the algorithms doubts their success, an intelligent learning system should adapt the difficulty level. Additionally, other methods for engagement can be incorporated to increase the motivation of students to finish an exercise. For example, providing hints for solving the solution. The constructed models can be personal, i.e., one model per learner, can be aggregated across clusters of students with similar user behaviors, or can be grouped in terms of general class performance. The learning outcomes and performance metrics from T2.1 can be used as possible values which can be predicted in unseen data.
T3.2 – Computational adaptation of course material (UoM) [M6-M16]
The predictive power of the models developed under T3.1 can be used in the automatic adaptation of the course material. Task 3.2 will transfer the mature knowledge of the game industry regarding procedural content generation (PCG) in order to automatically (or semi-automatically) generate a personalized course structure of the course material in order to accommodate the predicted learning performance of the user. Following the specifications from T1.3 regarding the learning outcomes and educators’ methodologies, the appropriate generative algorithms will be devised in order to (a) adapt the content in order to better assist the learning of at-risk (or over-competent) learners, while (b) still allowing the educator, when needed, to control the type of content presented and its order. The final algorithms are expected to obey educator-specific constraints on the structure of the virtual lab, while having the freedom to provide learners with content which is predicted as most likely to increase their engagement and learning potential. The predictive model will be updated based on the responses of learners to such personalized course material, leading to a real-time adaptive system which adapts itself and the course material to its users. In order to select appropriate content to show to learners, the algorithms developed under T3.2 will observe usage patterns of past learners with a similar usage behavior (as developed under T3.1) and how they engaged with course material; the material which resulted in the best learning outcomes (as specified by the learning analytics developed under T2.2) will be selected for current learners. As current learners progress in the courses, they become past learners and appropriately update the model of appropriate course material. This positive feedback loop allows the system to continuously learn and improve the delivery of the course material to better fit the learning processes of current learners.
UoM will lead the activities of this WP towards delivering learning content adapted to the specific needs of individual learners by consulting the predictions of machine learning algorithms.