LAUSR

Temporal graph networks are powerful tools for solving the cold-start problem in sequential recommender systems. However, graph models are susceptible to feedback loops and data distribution shifts. The paper proposes a simple yet efficient graph-based exploration method for the mitigation of the issues above. It adopts the counter-based state exploration from reinforcement learning to the bipartite graph domain. We suggest an approach that biases model predictions using Rooted PageRank towards locally unexplored items. The method shows competitive quality on the popular recommender systems benchmarks. We, also, provide an extensive qualitative analysis of experiment results and recommendations for our method production applications.