cripac-dig/micro

MIRCO

This is the code for the Paper: Latent Structure Mining with Contrastive Modality Fusion for Multimedia Recommendation.

Usage

Dataset Preparation

• Download 5-core reviews data, meta data, and image features from Amazon product dataset. Put data into the directory data/meta-data/.

• Install sentence-transformers and download pretrained models to extract textual features. Unzip pretrained model into the directory sentence-transformers/:

  ├─ data/: 
      ├── sports/
      	├── meta-data/
      		├── image_features_Sports_and_Outdoors.b
      		├── meta-Sports_and_Outdoors.json.gz
      		├── reviews_Sports_and_Outdoors_5.json.gz
      ├── sentence-transformers/
          	├── stsb-roberta-large
  

• Run python build_data.py to preprocess data.

• Run python cold_start.py to build cold-start data.

• We provide processed data Baidu Yun (access code: m37q), Google Drive.

Quick start

Start training and inference as:

cd codes
python main.py --dataset {DATASET}

For cold-start settings:

python main.py --dataset {DATASET} --core 0 --verbose 1 --lr 1e-5

Requirements

• Python 3.6
• torch==1.5.0
• scikit-learn==0.24.2
• torch-scatter==2.0.8

Citation

Please cite our paper if you use the code:

@article{zhang2022latent,
  title={Latent structure mining with contrastive modality fusion for multimedia recommendation},
  author={Zhang, Jinghao and Zhu, Yanqiao and Liu, Qiang and Zhang, Mengqi and Wu, Shu and Wang, Liang},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2022},
  publisher={IEEE}
}

Acknowledge

The structure of this code is largely based on LightGCN. Thank for their work.

micro's Issues

您好！代码和论文中的公式无法一一对应？

作者，您好！
我最近在复现您的工作MICRO。但是，当我阅读代码时，发现您的部分代码和论文中的公式无法对应上。例如：论文中的公式（11）~（13），在挖掘Modality-Specific信息时，我没有找到代码中对应的部分。另外，对于预测函数（公式17）和用户向量的融合（公式16）以及第二个注意力机制（公式14~15），均没有找到代码中对应的位置。
请问，这一部分代码在源码中的何处？

I'm struggling with the model reproduction.

Hi.

Fist of all, thank you for you guys opened this repository.
I learned alot with the paper and this code.

Before I try to injection my idea in this code, I wanted to reproduce the orignal model performance as you mentioned in the paper. But I've been being struggling with it. So, If you don't mind plz give me some advise about reproducing the model.

Here is what my experiment setting is.

• Using the preprocessed data shared on your google drive.

• Dataset sports

• n_users 35598

• n_items 18357

• n_interactions 268244

---

• lr 1e-5
• embed_size 64
• layers 2
• lambda_coeff=0.9
• cf_model 'lightgcn'
• mess_dropout=[0.1, 0.1]
• regs=[1e-4, 1e-4, 1e-2]
• sparse=1
• loss_ratio=0.03
• norm_type='sym'

---

This is my best score i got

Best regards.
Hoyoon.

请求一份baseline的代码

您好！我自己修改了在亚马逊和mmgcn的模型的数据集上grcn和mmgcn的模型的准确率要计算论文中的数据，请不要让你过的可以得到用于亚马逊数据集推荐的gcn和mmgcn的代码？

The questions about processing datasets.

Hi, authors.
Utilizing the processed datasets can reproduce the results reported in your paper.
However, when I rebuild needed datasets based on the raw data (e.g, reviews_Sports_and_Outdoors_5) by released codes (i.e., build_data.py), the performance is reduced severely.
Therefore, can you provide the remaining files like 'user-item-dict.json', 'user_list.txt', and 'item_list.txt' of the three datasets?