Implementation in TensorFlow 1.x of Contrastive Learning for Debiased Candidate Generation in Large-Scale Recommender Systems, KDD 2021 .
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.
multi_interest_clrec contains the implementation of the multi-interest sequence encoder, which can produce multiple vectors of a user for capturing the user's diverse interests.
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.