search

fengzhu1 / GA-DTCDR

This is the model in "A Graphical and Attentional Framework for Dual-Target Cross-Domain Recommendation" (IJCAI2020). GA-DTCDR is an optimized model for DTCDR ("DTCDR: A Framework for Dual-Target Cross-Domain Recommendation" in CIKM2019).
GNU General Public License v3.0
67 stars 18 forks source link

Hyperparameter setting #2

Closed lelele2001 closed 2 years ago

lelele2001 commented 2 years ago

When I used the hyperparameters provided in the source code or the paper and node2vec model provided in the source code, I did not achieve the experimental results in the paper. Sometimes, HR and NDCG are 1 for relatively small KSIZE,which is may caused by RELU function, so can I know the hyperparameter of getting the best result,please?

fengzhu1 commented 2 years ago

As introduced in the last sentence of "Result 3: Impact of Embedding Dimension k (for Q3)" of the paper, when KSIZE is 64, our model can achieve the best performance. This is because when the KSIZE is too small, the MLP will be too deep for such an embedding. If you want to choose a small KSIZE, you should adjust the depth and structure of the MLP. In fact, after the paper's publication, we are continuing to optimize this model and attempt to solve the gradient disappear problem when the KSIZE is too small. We will release a new optimized model in the near future for multi-target CDR scenarios.