A Theoretical Analysis of Two-Stage Recommendation for Cold-Start Collaborative Filtering
This work addresses the challenge of recommending to new users or items with limited data, though it appears incremental as it builds on existing POMDP frameworks.
The paper tackles the cold-start collaborative filtering problem by proposing a two-stage recommendation approach that balances exploitation and exploration, with initial results on synthetic data and MovieLens 100K showing performance gains.
In this paper, we present a theoretical framework for tackling the cold-start collaborative filtering problem, where unknown targets (items or users) keep coming to the system, and there is a limited number of resources (users or items) that can be allocated and related to them. The solution requires a trade-off between exploitation and exploration as with the limited recommendation opportunities, we need to, on one hand, allocate the most relevant resources right away, but, on the other hand, it is also necessary to allocate resources that are useful for learning the target's properties in order to recommend more relevant ones in the future. In this paper, we study a simple two-stage recommendation combining a sequential and a batch solution together. We first model the problem with the partially observable Markov decision process (POMDP) and provide an exact solution. Then, through an in-depth analysis over the POMDP value iteration solution, we identify that an exact solution can be abstracted as selecting resources that are not only highly relevant to the target according to the initial-stage information, but also highly correlated, either positively or negatively, with other potential resources for the next stage. With this finding, we propose an approximate solution to ease the intractability of the exact solution. Our initial results on synthetic data and the Movie Lens 100K dataset confirm the performance gains of our theoretical development and analysis.