Biases and Debiasing Strategies for Temporally Evolving Graph Recommendation Systems
Siddhant Saxena, Srikanta Bedathur, Shubham Gupta
Abstract
Research Report
Siddhant_IKDD_Report.pdf
Temporal Graph Recommendation Engine
In order to leverage the temporal evolution of the graph, so that the recommendation system compares and optimizes on the latest user preferences as underlying feature vectors, in this work we leverage Temporal Graph Networks (TGN) architecture as our baseline to temporal graph learning.
"diffDebias" Architecture
In this approach for the debiasing of the recommendation system, we formalize an end-to-end differentiable approach that can be integrated into the temporally evolving graph as a recommendation engine,. In this architecture, we propose a disentanglement-inspired approach to redistribute the observed link prediction distribution to ideal distribution, as we claim that the exposure mechanism of the recommendation engine becomes a source of biased recommendations, the implicit biases generate a distribution of the frequencies between user-item nodes which differ from the actual ideal distribution, which then gets amplified in the feedback loop.
Using this architecture we aim to remap the biased distribution to an ideal one, by exploiting the latent representation of the graph and propose an encoder-decoder-based debiasing framework.
Experimental Setup
