KDD 2022 | how far are we from the general pre training recommendation model? Universal sequence representation learning model unisrec for recommender system

The current research on sequence recommendation focuses on developing more efficient sequence representation learning （SRL） Model . Most of the existing methods are explicit to the goods ID Modeling the sequence , However, these models are difficult to migrate to new recommended scenarios , Such as new fields or platforms . in order to Solve modeling commodity ID constraints , This paper presents a new SRL Method UniSRec. Concrete ,UniSRec Use the text information of commodities to learn the common representation that can be migrated to different recommendation scenarios . In order to learn General merchandise means , We design a hybrid expert network based on parameter whitening （MoE） Enhanced commodity coding architecture ; In order to learn Universal sequence representation , We designed two optimization objectives based on contrastive learning , In the pre training phase, the sequences of multiple domains are sampled / Commodity as a negative example . The general sequence representation model after pre training can efficiently migrate parameters to new domains or platforms . A large number of experiments constructed on real data sets have verified UniSRec The effect of . Special , When hold Amazon Pre trained on dataset UniSRec Model migration to a new platform （ A British E-commerce ） when , You can also observe an increase in the effect , The strong mobility of the general sequence representation learning method proposed in this paper is verified .

Thesis title ： Towards Universal Sequence Representation Learning for Recommender Systems

Thesis download address ： 

https://arxiv.org/pdf/2206.05941

Thesis open source code ： 

https://github.com/RUCAIBox/UniSRec