REMIT: Reinforced Multi-Interest Transfer for Cross-Domain Recommendation

Authors: Caiqi Sun, Jiewei Gu, Binbin Hu, Xin Dong, Hai Li, Lei Cheng, Linjian Mo

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experimental results on large real-world datasets demonstrate the superior performance and compatibility of REMIT.
Researcher Affiliation Collaboration 1 Ant Group 2 School of Data Science, Fudan University {caiqi.scq, gujiewei.gjw, bin.hbb, zhaoxin.dx, tianshu.lh, lei.chenglei, linyi01}@antgroup.com
Pseudocode Yes Algorithm 1: Overall Training Procedure of REMIT
Open Source Code Yes REMIT 2Code is available at https://github.com/mufusu21/REMIT
Open Datasets Yes Following most existing methods (Zhu et al. 2021, 2022), we test our algorithms on the Amazon review dataset1. Specifically, we pick 3 datasets out of 24, i.e., movies and tv (Movie), cds and vinyl (Music), and books (Book). Three cross-domain recommendation tasks are built upon these three datasets: Movie Music (Task 1), Book Movie (Task 2) and Book Movie (Task 3). ... 1http://jmcauley.ucsd.edu/data/amazon/
Dataset Splits No The paper describes how test users are defined by percentages (80%, 50%, 20%) and that other overlapping users are used for training, but it does not explicitly mention a separate validation split.
Hardware Specification No The paper only states "built based on the code repository of PTUPCDR using Py Torch and GPU" without specifying exact hardware details like GPU model, CPU, or memory.
Software Dependencies No The paper mentions "Py Torch" but does not provide a specific version number or other software dependencies with their versions.
Experiment Setup Yes For both MIT and IS in each task and method, the initial learning rate for the Adam (Kingma and Ba 2014) optimizer is tuned by grid searches within {0.001, 0.005, 0.01, 0.02, 0.1}. The epoch number is tuned by grid searchs between 5 and 15. In addition, we set the dimension of embeddings as 10 and the batch size as 512. The meta-path set is {uiu, uiciu, uibiu}, which represents users interests of item, category and brand. The hidden size of meta network in MIT and policy network in IS is set to 50 and 25.