Fair Multimodal Recommendation

This repository includes the implementation for paper Causality-Inspired Fair Representation Learning for Multimodal Recommendation.

Datasets

The preprocessed MovieLens-1M dataset are already provided in the ./data/ml1m folder. The proprocessed data of MicroLens dataset could be downloaded from MicroLens-Fairness.

Environments

The experimental environment is Python 3.10.11. We can first create and activate a new Anaconda environment for Python 3.10.11:

> conda create -n FMMRec python=3.10.11
> conda activate FMMRec

Then install all the required packages by using the command:

> pip install -r ./requirements.txt

Usage

The used disentangled modal embeddings are already contained in the ./data/[dataset]/ folder. To run the disentanglement learning, for example, we could run the following code for visual modality on the MicroLens dataset:

> python BMMF_runner.py --dataset microlens --modality v --gpu_id 0 --epochs 100

For the MovieLens dataset, we can run the code of the assembly of FMMRec fairness method on LATTICE recommendation model by running this command:

> cd ./src/
> nohup python -u main.py --fairness_model BFMMR --knn_k_uugraph 10 --filter_mode shared --prompt_mode concat --recommendation_model LATTICE --dataset ml1m --d_steps 10 --gpu_id 1 > MovieLens.out 2>&1 &

For the MicroLens dataset, we can run the code of the assembly of FMMRec fairness method on DRAGON recommendation model by running this command:

> cd ./src/
> nohup python -u main.py --fairness_model BFMMR --knn_k_uugraph 7 --filter_mode shared --prompt_mode concat --recommendation_model DRAGON --dataset microlens --d_steps 10 --gpu_id 0 > MicroLens.out 2>&1 &

Acknowledgement

The code of this repository is implemented based on the multimodal recommendation framework at MMRec.