Comparing K-Means and Latent Profile Analysis for Profiling Online Poker Players from Behavioral Features

Sammanfattning

This thesis investigates the use of unsupervised learning methods for profiling online poker players based on continuous behavioural features. A comparison is made between distance-based clustering (K-means and K-means++) and model-based clustering (LPA) when applied to Pot-Limit Omaha hand-history data. The methods are evaluated through model-selection criteria, robustness across repeated runs, interpretability of the resulting profiles, and their alignment with external profitability measures excluded from the clustering process. The results show that all methods were able to separate winning and losing players to some extent, and both approaches produced robust profiles, but K-means and especially K-means++ produced more interpretable and meaningful player profiles than LPA. The K-means-based solutions identified clear strategic archetypes, while the most profitable clusters were characterised by controlled pre-flop play and higher postflop aggression. LPA showed strong assignment certainty but struggled to isolate extreme losing behaviour, suggesting that probabilistic mixture models may be less suitable for this dataset.