MMRA：Predicting Micro-video Popularity via Multi-modal Retrieval Augmentation

In this repository, we provide source code of MMRA framework for reproductivity 🚀🚀🚀. Paper doi: https://dl.acm.org/doi/abs/10.1145/3626772.3657929.

Dataset

We conducted extensive experiments on a real-world micro-video dataset.

MicroLens

MicroLens consists of 19,738 unique micro-videos posted from 100000 users, sourced from various popular online video platforms. This is the link of the original dataset: https://github.com/westlake-repl/MicroLens

Usage

Requirements

The code was tested with python 3.8.18, pytorch 2.1.1, cudatookkit 12.1. Install the dependencies via Anaconda:

# create virtual environment

conda create --name MMRA python=3.8

# activate environment

conda activate MMRA

# install pytorch & cudatoolkit

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

# install other requirements

pip install pandas

pip install numpy

pip install scipy

pip install scikit-learn

pip install tqdm

pip install transformers

Run code

To train MMRA framework on a dataset, run the following command:

python train.py --device <device>  --dataset_path <dataset_path> --dataset_id
<dataset_id> --metric <metric>  --save <save>  --epochs <epochs>  --batch_size <batch_size>  --early_stop_turns <early_stop_turns>  --seed <random_seed>  --loss <loss>  --optim <optimizer>  --lr <learning_rate>  --model_id <MMRA>  --feature_num <feature_num>  --feature_dim <feature_dim>  --label_dim <label_dim>  --alpha <alpha>  --num_of_retrieved_items <num_of_retrieved_items>  --frame_num <frame_num>  

To evaluate MMRA framework on a dataset, run the following command:

python test.py --device <device>  --dataset_path <dataset_path> --dataset_id
<dataset_id> --metric <metric>  --save <save>  --batch_size <batch_size>  --seed <random_seed>  --model_id <MMRA>  --feature_num <feature_num>  --feature_dim <feature_dim>  --label_dim <label_dim>  --alpha <alpha>  --num_of_retrieved_items <num_of_retrieved_items>  --frame_num <frame_num>   --model_path <model_path>

To retrieve similar items, run the following command:

python retriever.py

The detailed descriptions about the important arguments are as following:

Parameter name	Type	Description
dataset_id	str	The dataset id, default is "MicroLens-100k".
save	str	Directory to save results.
loss	str	Loss function used in training, default is "MSE".
feature_num	int	Number of kinds of features, default is 2.
feature_dim	int	The dimension of embedding of features.
label_dim	int	The dimension of label, default is 1.
alpha	float	Positive Negative attention balanced parameter alpha, default is 0.6.
num_of_retrieved_items	int	Number of retrieved items, hyper-parameter.
frame_num	int	Number of frames, hyper-parameter.
lr	float	Learning rate, in this work, its default valud is 1e-5 .

Directory Structure

The code directory structure is shown as follows:

MMRA
├── baselines  # source code of baselines rewrited in our work
│   ├── CBAN  
│   ├── Contextual_LSTM 
│   ├── HMMVED 
│   ├── Hyfea
│   ├── MASSL 
│   ├── MFTM 
│   ├── SVR 
│   └── TMALL 
├── data  # source code of data-preprocessing and retrieval preprocessing
│   ├── MicroLens-100k
│	│	 ├── data_preprocess
│	│	 ├── retrieval_preprocess
├── dataloader  # dataloader for training, validation and test
│   ├── MicroLens
│	│	 ├── dataset.py 
├── feature_engineering  # feature engineering source code
│   ├── textual_engineering.py 
│   ├── visual_engineering.py 
│   └── match_image_feature_to_video_item.py    
├── model  # code of MMRA
│   ├── MicroLens
│	│	 ├── MMRA.py 
├── Retriever.py # code of Retriever used in MMRA
├── train.py  # entry for model training,validation
├── test.py  # entry for model test
└── README.md  # This file

Citation

If you find the code useful for your research, please consider citing

@inproceedings{zhong2024predicting,
  title = {Predicting Micro-video Popularity via Multi-modal Retrieval Augmentation},
  author = {Zhong, Ting and Lang, Jian and Zhang, Yifan and Cheng, Zhangtao and Zhang, Kunpeng and Zhou, Fan},
  booktitle = {Proceedings of the 47th ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2024}
}

If you have any questions about this code or the paper, feel free to contact us: jian_lang@std.uestc.edu.cn