Predicting Facebook Ad Campaign Performance Using Few-Shot Learning with Large Language Models

Mgr. Shahadat Hussain

Diplomová práce

Abstract:

Predicting the success of digital advertising campaigns is challenging, especially when historical data are scarce. This study investigates the ability of three large language models, DeepSeek v3.2, GPT OSS 120b, and Qwen3.5, to predict Facebook ad campaign outcomes using few- shot learning. A cleaned and preprocessed dataset of 761 campaigns was used. In this thesis, success is defined as a binary outcome based on three advertising performance criteria: click-through rate, cost per click, and conversion rate. A campaign is labelled successful if it satis- fies at least two of these three criteria. The final dataset contains 167 successful campaigns and 594 unsuccessful campaigns, producing a success rate of 21.94 percent. The models were evaluated under zero-shot, one-shot, three- shot, and five-shot conditions using chain-of-thought prompting and similarity-based example selection. The latest experimental results show that Qwen3.5 achieves the strongest overall balanced perfor- mance in the five-shot setting, with 94.77 percent accuracy, 95.59 per- cent balanced accuracy, and an MCC of 0.8626. GPT OSS 120b also achieves strong five-shot performance, with 94.77 percent accuracy, 92.44 percent balanced accuracy, and an MCC of 0.8487. DeepSeek v3.2 performs strongly in the zero-shot setting, with 88.89 percent accuracy and 92.86 percent balanced accuracy, but does not improve consistently with additional examples. These findings show that LLMs can reason over structured tabular marketing data when prompts are carefully designed and evaluated with imbalance-aware metrics.

Keywords

Few-Shot Learning, Large Language Models, Facebook Ads, Campaign Performance Prediction, Prompt Engineering, Tabular Data.

Jazyk práce: angličtina

Datum vytvoření / odevzdání či podání práce: 19. 5. 2026

Obhajoba závěrečné práce

• Obhajoba proběhla 18. 6. 2026
• Vedoucí: Mgr. Marek Grác, Ph.D.
• Oponent: RNDr. Adam Rambousek, Ph.D.

Masarykova univerzita

Fakulta informatiky

Magisterský studijní program / obor:
Software Systems and Services Management / Smart Service Design and Systems Thinking