Knowledge tracing (KT) is an educational binary prediction task that predicts future performance of students based on their given information of the students and the questions. KT processes data collected from web-based Intelligent Tutoring Systems (...
Knowledge tracing (KT) is an educational binary prediction task that predicts future performance of students based on their given information of the students and the questions. KT processes data collected from web-based Intelligent Tutoring Systems (ITSs) or offline tests to predict whether future students will answer questions correctly. Previous works on KT have focused mainly on making performance predictions based on sufficient past interaction records from existing students. Also, recent attempts have been made to build KT systems in few-shot settings where little data of the target student is provided in advance. In these tasks, the predictions are mainly done based on the traditional cold-start settings, where minimal data is presented. However, there were relatively fewer attempts to increase KT performance in a situation where the entire interaction records between students and questions are not given, which is defined as system cold-start settings. Meanwhile, there have been attempts to utilize Large Language Models (LLMs) as information providers, inference machines, and probability machines to empower recommendations. In this context, we present System Cold-Start Knowledge Tracing (SCSKT) framework that extends the KT task to system cold-start settings, utilizing the inherent abilities of LLMs. First, an LLM as an information provider takes the information of a student and a question as input and performs a feature selection process, preserving important unit features by ranking. Then, an LLM as an inference machine takes the unit features as input and performs a feature rearrangement based on pedagogical assumptions. Lastly, an LLM as a probability machine takes the unit features as input and performs a binary prediction on whether the student will solve the given question under given conditions. To evaluate our framework, we utilize three real-world datasets from web-based online ITS platforms and government-hosted offline tests for experiments. The experiment shows increased performance of our framework compared to the existing baselines. To the best of our knowledge, SCSKT is the first knowledge tracing framework that utilizes LLM to enhance performance in system cold-start settings, providing smooth first interactions between the various types of educational questions and the students.