Awesome
Paper Accepted to WSDM'22
-
Title: Linear, or Non-Linear, That is the Question!
How to Use
docker pull pytorch/pytorch
docker run --gpus all -it --rm --privileged -v {local_path}:/HMLET pytorch/pytorch bash -c "pip install pandas && pip install scipy && pip install sklearn && pip install tensorboardX && pip install openpyxl && cd /HMLET && {train_model_command}"
python train.py --dataset {dataset_name} --model {model_variants}
Methods Proposal Background and Purpose
-
Which embedding propagation (linear & non-linear) is more appropriate to recommender systems?
Methods
-
HMLET (Hybrid-Method-of-Linear-and-non-linEar-collaborative-filTering-method)
-
Dynamically selecting the best propagation method for each node in a layer using gating networks.
-
Four variants of HMLET: HMLET (End), HMLET (Middle), HMLET (Front), HMLET (All)
-
Four variants of HMLET in terms of the location of the non-linear propagation.
<p align="center">
<img src="https://user-images.githubusercontent.com/52263269/141878827-d40a2844-8fad-4d75-aae3-0f693bb1034c.png" width="550px" height="350px"></img>
</p>
-
HMLET (End)
-
HMLET (End) shows best performance among these variants
-
Focusing on gating in the third and fourth layers
-
The detailed workflow of HMLET (End)
<p align="center">
<img src="https://user-images.githubusercontent.com/52263269/141666368-71bff1c9-f4a4-4ffd-b6ca-f0ecbdf5f845.png" width="1100px" height="350px"></img>
</p>