BiasedUserHistorySynthesis

Source for Biased User History Synthesis for Personalized Long Tail Item Recommendation

Create the conda environment

conda create --name ENV_NAME --file conda_requirements.txt python=3.7
conda activate ENV_NAME
conda install pip
python3 -m pip install -r requirements.txt

Downloading the datasets

We use two public benchmark datasets: MovieLens1m and BookCrossing. Download the datasets into a datasets subdirectory as follows: PATH/TO/DATASETS_DIR/ml-1m and PATH/TO/DATASETS/BookCrossing, and subsequently seting the following environment variable:

export DATASETS_DIR=PATH/TO/DATASETS

Reproducing Results

The default branch of this repository is called BaseModel. It contains the code for the Two Tower Neural Network Base Recommendation System. The other branch in this repository is called BiasedUserHistorySynthesis. It contains the code for BiasedUserHistorySynthesis built on top of a Two Tower Neural Network. The main results in the paper for the Base Model and all variants of BiasedUserHistorySynthesis can be reproduced by running the following shell scripts:

Dataset	Model	Command
MovieLens-1m	Base Two Tower Neural Network	bash ml1m_ttnn_basemodel.sh
MovieLens-1m	BUHS-Mean + TTNN	bash ml1m_ttnn_buhs_mean.sh
MovieLens-1m	BUHS-Attn + TTNN	bash ml1m_ttnn_buhs_attn.sh
MovieLens-1m	BUHS-GRU + TTNN	bash ml1m_ttnn_buhs_gru.sh
BookCrossing	Base Two Tower Neural Network	bash bx_ttnn_basemodel.sh
BookCrossing	BUHS-Mean + TTNN	bash bx_ttnn_buhs_mean.sh
BookCrossing	BUHS-Mean + TTNN	bash bx_ttnn_buhs_attn.sh
BookCrossing	BUHS-Mean + TTNN	bash bx_ttnn_buhs_gru.sh