Implementation in TensorFlow 1.x of Contrastive Learning for Debiased Candidate Generation in Large-Scale Recommender Systems.

clrec_v1_local_sasrec is the version that we use for conducting the experiments on the public datasets, which runs on a single machine. Note that this version uses the same neural architecture as SASRec in order to investigate the effect of the contrastive loss itself.

clrec_v1_distributed is the version that we use for conducting the large-scale experiments in our real-world production environment, which runs on the distributed clusters provided by our company's infrastructure. This version uses the neural architecture described in the appendix of our paper.

The public datasets used in our paper are pre-processed and provided by SASRec.

Please find the description of the new dataset and its download link here. We could only release a sampled anonymized subset of our production environment data, while we conducted the experiments in our paper on the non-sampled large-scale data.