SPELLS — A Representation Learning Approach to Latent Power Prediction for Magic: The Gathering

Conneely, Sam (2025) SPELLS — A Representation Learning Approach to Latent Power Prediction for Magic: The Gathering. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (4MB) | Preview
Preview	PDF (Configuration Manual) Download (2MB) | Preview

Abstract

This thesis introduces SPELLS (Semi-Supervised Power Estimation and Label Learning System), a multimodal learning framework for estimating the latent power level of Magic: The Gathering cards. With no standardised power metrics and manual annotation of 26,795 cards being impractical, SPELLS uses a small human labelled subset (< 0.4% of cards) with multimodal embeddings derived from the oracle text, artwork, and structured metadata encoded on the cards. Anchor-graph label propagation generates pseudo-labels across the entire dataset, allowing for the training of regression models including a deep neural network, XGBoost and Support Vector Regressor.

Extensive preprocessing was done including the normalisation of oracle text as well as the generation of Sentence-BERT embeddings (384-dimensional) for keyword and type information, ResNet50 visual features (reduced to 256-dimensions via PCA), and engineered structural features. An active learning stage with human-in-the-loop correction was done iteratively to address outliers in the label propagation.

Evaluation on the manually labelled anchors shows that all models achieve Mean Absolute Error of < 1.0 on the 1-10 power scale. Further testing on cluster representative cards from the full manifold structure reveals that the DNN (R2 = 0.87, Spearman=0.97) and XGBoost (R2 = 0.83, MAE=0.19, Spearman=0.97) successfully learned the latent power, while Support Vector Regressor failed to maintain stable predictions on fused embeddings using all three modalities (MAE= 1.26 R2 = −0.28).

This work proves that semi-supervised multimodal learning can effectively estimate subjective latent features in complex game systems with minimal expert annotation. SPELLS proves a foundation for future research on format specific modelling and transferability to other trading card games and domains requiring subjective quality assessment from heterogenous features.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Jameel Syed, Muslim UNSPECIFIED
Uncontrolled Keywords:	Semi-supervised learning; multimodal fusion; label propagation; trading card games; Magic: The Gathering power estimation
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence G Geography. Anthropology. Recreation > GV Recreation Leisure > Games and Amusements
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	02 Sep 2026 08:56
Last Modified:	02 Sep 2026 08:56
URI:	https://norma.ncirl.ie/id/eprint/9752

Actions (login required)

View Item