41ST ANNUAL ACM SYMPOSIUM ON APPLIED COMPUTING, SAC 2026
Abstract
Decisions about study choice profoundly affect pupils' educational pathways and long-term career prospects, yet many still struggle to select options that align with their skills and interests. To support more personalized study choice, a serious game named "Kr8cht" was designed that elicits Flemish pupils' preferences through pairwise comparisons of activities. Each activity is linked to study domains and personality traits based on expert-annotated relevance scores. However, the current version of the game includes only a limited set of activities, restricting its ability to accommodate the diverse needs and preferences of pupils. To address this limitation, we extensively investigate multi-target regression (MTR) techniques to automatically predict domain scores for newly suggested activities, thereby reducing dependence on expert annotation. Our findings show that local and chain-based MTR methods trained on embeddings from frozen sentence-level transformers consistently outperform global MTR approaches and those trained on embeddings from static or frozen token-level transformer models. Furthermore, our experiments also reveal that augmenting the dataset with synthetic activities, which are pseudolabelled by a MTR model, increases performance with statistical significance. The source code for our experiments is available at Zenodo (https://doi.org/10.5281/zenodo.17834064).