This seminar includes step-by-step decomposition of jaywonchung's BERT4Rec implementation and comparing the ranking perfomance of this implementation with some baselines such as random sort, popular sort and some matrix factorizations. The results of these benchmarks are consistent with the results from the original paper BERT4Rec: Sequential Recommendation with BERT (Sun et al.).
The main goal of this seminar is to understand BERT4Rec mechanism by examining an effect of each pytorch operation on specific batch sample.
Video of the seminar with explanations: here