TEM Journal | 2026
Authors: Mitreva E.
DOI: 10.18421/TEM151-80
Journal: TEM Journal
Year: 2026
Publisher: UIKTEN - Association for Information Communication Technology Education and Science
Document Type: Article
Open Access: All Open Access; Gold Open Access
Cited by: 0
Classifying short text poses significant challenges, especially when only a small portion of the data is labeled. Although supervised learning methods can achieve high accuracy on labeled datasets, overfitting often occurs, leading to poor generalization on unseen data. The primary concern is the significant decline in model accuracy when transitioning from a small labeled dataset to a larger set of unlabeled data. To address this issue, the article explores a few strategies: semi-supervised learning, k-means clustering, and stacking of classifiers. Semi-supervised learning leverages both labeled and unlabeled data by iteratively generating pseudo-labels and assigning confidence-based weights, thus expanding the training set without amplifying noise, because greater weight is assigned to more reliable labels. K-means clustering is similarly employed to partition unlabeled data into clusters, and pseudo-labels are inferred via majority voting based on similarity to labeled examples. These augmented methods are shown to mitigate the decline in accuracy observed when transitioning from the small, labeled dataset to previously unseen, unlabeled data. Although stacking multiple classifiers and methods to mitigate imbalanced classes did not improve performance, there are notable mentions and the lack of improvement was likely due to uniformly propagating errors from models that already struggle on unseen samples. © 2026 Emanuela Mitreva; published by UIKTEN. This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 License.
semi-supervised learning; Supervised-learning; unsupervised-learning