E2Vec: Feature Embedding with Temporal Information for Analyzing Student Actions in E-Book Systems

A research team has introduced E2Vec, a feature representation method that encodes the timing of student actions in e-book systems to improve at-risk detection, according to a study presented at the 17th Educational Data Mining Conference [1]. The method, detailed in a paper led by Yuma Miyazaki and including co-author Valdemar Švábenský, treats operation logs and their time intervals as character sequences, then applies the fastText word-embedding algorithm to generate a vector for each student [1][3]. The dataset comprised 305 students across two years of computer science courses [1]. Conventional approaches to analyzing EventStream data — the sequence of interactions recorded by digital textbook platforms — have relied on statistical features such as operation-type counts or access frequencies [2]. Those features, the authors note, “lack temporal information that captures fine-grained differences in learning behaviors among different students” [2]. E2Vec addresses the gap by modeling both the order of operations and the intervals between them [4]. The architecture borrows concepts from natural language processing. Short learning activities — opening an e-book and flipping pages at certain intervals — are represented as “units,” analogous to words, while longer sequences of units form “actions,” analogous to sentences [4]. The fastText model then learns distributed representations of these units, so that similar units acquire similar vector representations [6]. In a subsequent study that adopted E2Vec for learner modeling, researchers described how the representations are aggregated using a method inspired by Bag of Visual Words, yielding 100-dimensional feature vectors per learner [5]. That later work omitted the L2-normalization step used in the original paper to retain information about the volume of actions a student generated [5]. The authors tested E2Vec on an at-risk prediction task, a standard benchmark in educational data mining, and reported a higher F1-score than models using only operation-count features [4]. The paper was published in the proceedings of EDM 2024 and is available on arXiv [3].