From Missing Data to Boltzmann Distributions and Time Dynamics: The Statistical Physics of Recommendation

The challenge of building a good recommendation system is deeply connected to missing data---unknown features and labels to suggest the most "valuable" items to the user. The mysterious properties of the power law distributions that generally arises out of recommender (and social systems in general) create skewed and long-tailed consumption patterns that are often still puzzling to many of us. Missing data and skewed distributions create not just accuracy and recall problems, but also capacity allocation problems, which are at the roots of recent debate on inclusiveness and responsibility. So how do we move forward in the face of these immense conceptual and practical issues? In our work, we have been asking ourselves ways to deriving insights from first principles and drawing inspiration from fields like statistical physics. Surprised, one might ask---what does the field of physics has to do with missing data in ranking and recommendations? As we all know, in the field of information systems, concepts like information entropy and probability have a rich intellectual history. This history is deeply connected to the greatest discoveries of science in the 19th century---statistical mechanics, thermodynamics, and specific concepts like thermal equilibrium. In this talk, I will take us on a journey connecting Boltzmann distribution and partition functions from statistical mechanics with importance weighting for learning better softmax functions, and then further to reinforcement learning, where we can plan better explorations using off-policy correction with policy gradient approaches. As I shall show, these techniques enable us to reason about missing data features, labels, and time dynamic patterns from our data.