Topology of preference networks for user–item recommendation.

Research · Plate VII

Topological Methods for Recommender Systems

User–item preference networks · cold-start · long-tail Three Master's students Updated MMXXVI

The user–item interaction graph: a bipartite network whose topology—connectedness, loops, voids in the preference space—carries signal that ordinary matrix-factorisation and embedding pipelines discard.

“Topology has been quiet about preferences for too long.”

A recent thread, with three Master's students at GWU, applies topological data analysis to recommender systems. The user–item interaction graph is a bipartite network with structure that ordinary matrix-factorisation and embedding pipelines politely throw away: connectedness, loops, voids in the preference space, the cohomology of how taste clusters and crosses over.

The work asks how persistence-based and Euler-characteristic descriptors of that network can sharpen recommendation, especially in the cold-start and long-tail regimes where conventional embeddings are starved of co-rating data and reduce, in practice, to popularity.

Master’s students

Filed under Co-authors The George Washington University

• Alexander D. Silberman, M.S. Data Science, GWU
• Chinaza Belolisa, M.S. Data Science, GWU
• Madeline Bumpus, M.S. Data Science, GWU

❦

Back to the research overview, or read about the closely related work on topological kernels for machine learning.