Context-Dependent Clustering for Exploration-Exploitation in Recommendation Systems
- Speaker: Professor Claudio Gentile, Insubria University, Italy
- Date: Wednesday, 10 August 2016 from 14:00 to 15:00
- Location: Room 151
Bandit algorithms have become a standard tool for facing the the so-called exploration-exploitation dilemma that naturally arises in learning problems with partial information.
Recommendation systems are one of the best applications where such algorithmic solutions may be of relevance.
In many cases, these applications have social components (either explicit or hidden), whose integration in the bandit
algorithms could lead to a significant performance increase.
For instance, we may want to serve content to a group of users by taking advantage of an underlying network of social relationships among them.
In typical scenarios like social networks, it is often possible to single out a few groups or communities made up of users sharing similar interests.
Such communities are not static over time and, more often than not, are clustered around specific content types, so that a given set of users can in fact host a multiplex of interdependent communities depending on specific content items or group of items.
We call this multiplex of interdependent clusterings a context-dependent clustering.
Typically, these clusterings are unknown to the learning agent, and have to be inferred on the fly based on the observed data.
In this talk, I'll review recent research activity in the framework of stochastic bandit algorithms where users are structured in (unknown) context-dependent clusters.
I will briefly present algorithms, associated theoretical analyses, and empirical evidence on real-world data.
Joint work with: G. Zappella, S. Li, and A. Karatzoglou.