Work

Asmodee Recommendation System

Case Study 14 March 2025
Data Science
Recommender System
Graph Theory
Product Strategy

A four-month deep dive into 200M rows of Board Game Arena data to understand what makes players stay — and what game to recommend next.

Network graph and clustering interface from the Asmodee project

Description

As part of Asmodee’s Business Deep Dive, we worked on a massive dataset of over 200 million rows from Board Game Arena to analyze player behavior and game interactions. With Anna Spira, we combined graph modeling, clustering, and recommendation techniques to design a personalized, data-driven gaming experience.

Key Features

  • Game Graph Visualization: We built an interactive graph of 1,100+ games, where edges were classified using cosine similarity as complementary or competitive. The graph was enriched via the Board Game Geek API and allowed users to dynamically add new games.

  • Player Segmentation: Using hierarchical clustering and Gaussian Mixture Models, we segmented players into six profiles based on activity, engagement, and reputation metrics.

  • Recommender Engine: We developed a custom algorithm combining graph proximity, user segment, and business rules to suggest relevant games for each type of player.

Technologies Used

  • Python: Core language for all data processing and pipeline logic.
  • Pandas: Data wrangling on 200M+ rows of gameplay data.
  • NumPy: Numerical operations and cosine similarity computations.
  • NetworkX: Graph creation and game relationship modeling.
  • Scikit-learn: GMM, hierarchical clustering, and preprocessing.
  • Plotly: Interactive visualizations for the recommendation interface.

My Role

This was a four-month case study with Anna Spira, and I owned the technical backbone of the project. I built the end-to-end game graph pipeline — fetching data from the Board Game Geek API, computing cosine similarities between 1,100+ games, and turning the result into an interactive Plotly visualization. I also co-designed the player segmentation and the recommender rules that tied everything together.

What I Learned

Working with 200M rows taught me that scale changes everything: a clever algorithm means little if the preprocessing pipeline is fragile. I also learned that product teams care more about interpretability than raw accuracy — a recommendation only matters if the business can explain why it was made.

Outcome

Our work was selected as a finalist and highlighted for its ability to transform complex behavioral data into a clear, actionable product strategy. The interactive graph and recommender system gave Asmodee tools to better engage and retain their player base.