Explainability of Deep RL algorithms using graph networks and layer-wise relevance propagation.

In this project we trained personalized transformer models for news recommendation using adapters [similar to (IA)^3]. With layerwise relevancy propagation, we try to explain the recommendation to the user. Using a web interface and displaying word clouds, the user can be assigned to a “filter bubble”. This allows users to reflect on their behavior

An explainability model that can be applied to BERT-based Turkish sentiment analysis models has been developed and its performance has been compared with model spesific Layer-wise relevance propogation expailanbility model of Hila Chefer.