Prompt-Based Continual Learning on Telecom Tabular Data
2025 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE credits
Student thesis
Abstract [en]
Continual learning addresses the challenge of adapting machine learning models to evolving data while avoiding full retraining and preserving previously acquired knowledge. While prompt-based continual learning has shown promise in natural language and computer vision tasks, its effectiveness on structured tabular data is not well studied. This thesis explores whether instance-wise prompting can improve generalization and reduce forgetting in structured telemetry data collected from a real 5G network.We adapt a transformer-based model designed for tabular regression and integrate it with a continual learning mechanism that selects prompts based on each individual input. The approach is evaluated in a domain-incremental setting, where each domain corresponds to a distinct network deployment scenario. These domains share the same prediction target but differ in feature distribution, reflecting real-world variations across devices and locations.Experimental results show that the prompt-enhanced model consistently outperforms a strong non-prompted baseline in terms of prediction accuracy and stability across domains. The improvements are especially notable in domains that exhibit significant distributional shifts. The model is able to generalize better to new domains while maintaining performance on earlier ones, demonstrating reduced forgetting and improved temporal consistency. This suggests that prompt-based methods can offer an effective solution for continual learning on structured data, supporting efficient model adaptation in dynamic environments without sacrificing previously learned knowledge.
Place, publisher, year, edition, pages
2025. , p. 40
Keywords [en]
continual learning, prompt learning, tabular transformer, telecom data
National Category
Artificial Intelligence
Identifiers
URN: urn:nbn:se:uu:diva-562837OAI: oai:DiVA.org:uu-562837DiVA, id: diva2:1979992
External cooperation
Ericsson AB
Subject / course
Statistics
Educational program
Master Programme in Statistics
Presentation
2025-06-04, B159, Ekonomikum, Kyrkogårdsgatan 10, 753 12, Uppsala, 12:46 (English)
Supervisors
Examiners
Note
The experimental data used in this thesis were provided by Ericsson AB.
2025-07-012025-07-012026-07-01Bibliographically approved