A Toolkit for Neural Review-based Recommendation models with Pytorch. 基于评论文本的深度推荐系统模型库 (Pytorch)

Update: 2021.03.29

Add a branch (PL) to use PyTorch Lightning to wrap the framework for further distributed training.

git clone https://github.com/ShomyLiu/Neu-Review-Rec.git
git checkout pl

And the usage:

python3 pl_main.py run --use_ddp=True --gpu_id=2
# indicating that using ddp mode for distributed training with 2 gpus. refer `config/config.py`.

Neural Review-based Recommendaton

In this repository, we reimplement some important review-based recommendation models, and provide an extensible framework NRRec with Pytorch. Researchers can implement their own methodss easily in our framework (just in models folder).

Introduction to Review-based Recommendaiton

E-commerce platforms allow users to post their reviews towards products, and the reviews may contain the opinions of users and the features of the items. Hence, many works start to utilize the reviews to model user preference and item features. Traditional methods always use topic modelling technology to capture the semantic informtion. Recently, many deep learning based methods are proposed, such as DeepCoNN, D-Attn etc, which use the neural networks, attention mechanism to learn representations of users and items more comprehensively.

More details please refer to my blog.

Methods

Note: since each user and each item would have multiple reviews, we categorize the existing methods into two kinds:

document-level methods: concatenate all the reviews into a long document, and then learn representations from the doc, we denote as Doc feature.
review-level methods: model each review seperately and then aggregate all reviews together as the user item latent feature.

Besides, the rating feature of users and items (i.e., ID embedding) is usefule when there are few reviews. So there would be three features in all (i.e., document-level, review-level, ID).

We plan to follow the state-of-art review-based recommendation methods and involve them into this repo, the baseline methods are listed here:

Method	Feature	Status
DeepCoNN(WSDM'17)	Doc	✓
D-Attn(RecSys'17)	Doc	✓
ANR(CIKM'18)	Doc, ID	☒
NARRE(WWW'18)	Review, ID	✓
MPCN(KDD'18)	Review	✓
TARMF(WWW'18)	Review, ID	☒
CARL(TOIS'19)	Doc, ID	☒
CARP(SIGIR'19)	Doc, ID	☒
DAML(KDD'19)	Doc, ID	✓

We will release the rest unfinished baseline methods later.

References

Zheng L, Noroozi V, Yu P S. Joint deep modeling of users and items using reviews for recommendation[C]//Proceedings of the Tenth ACM International Conference on Web Search and Data Mining. ACM, 2017: 425-434.
Sungyong Seo, Jing Huang, Hao Yang, and Yan Liu. 2017. Interpretable Convolutional Neural Networks with Dual Local and Global Attention for Review Rating Prediction. In Proceedings ofthe Eleventh ACMConference on Recommender Systems.
Chin J Y, Zhao K, Joty S, et al. ANR: Aspect-based neural recommender[C]//Proceedings of the 27th ACM International Conference on Information and Knowledge Management. 2018: 147-156.
Chen C, Zhang M, Liu Y, et al. Neural attentional rating regression with review-level explanations[C]//Proceedings of the 2018 World Wide Web Conference. 2018: 1583-1592.
Tay Y, Luu A T, Hui S C. Multi-pointer co-attention networks for recommendation[C]//Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2018: 2309-2318.
Lu Y, Dong R, Smyth B. Coevolutionary recommendation model: Mutual learning between ratings and reviews[C]//Proceedings of the 2018 World Wide Web Conference. 2018: 773-782.
Wu L, Quan C, Li C, et al. A context-aware user-item representation learning for item recommendation[J]. ACM Transactions on Information Systems (TOIS), 2019, 37(2): 1-29.
Li C, Quan C, Peng L, et al. A capsule network for recommendation and explaining what you like and dislike[C]//Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval. 2019: 275-284.
Liu D, Li J, Du B, et al. Daml: Dual attention mutual learning between ratings and reviews for item recommendation[C]//Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2019: 344-352.

Usage

Requirements

Python >= 3.6
Pytorch >= 1.0
fire: commend line parameters (in config/config.py)
numpy, gensim etc.

Use the code

Preprocessing the origin Amazon or Yelp dataset via pro_data/data_pro.py, then some npy files will be generated in dataset/, including train, val, and test datset.
```
cd pro_data
python3 data_pro.py Digital_Music_5.json
# details in data_pro.py (e.g., the pretrained word2vec.bin path)
```
Train the model. Take DeepCoNN and NARRE as examples, the command lines can be customized:
```
python3 main.py train --model=DeepCoNN --num_fea=1 --output=fm
python3 main.py train --model=NARRE --num_fea=2 --output=lfm
```
Note that the num_fea (1,2,3) corresponds how many features used in the methods, (ID feature, Review-level and Doc-level denoted above).
Test the model using the saved pth file in checkpoints in the test datase: for example:
```
python3 main.py test --pth_path="./checkpoints/THE_PTH_PATH" --model=DeepCoNN
```

An output sample:

loading train data
loading val data
train data: 51764; test data: 6471
start training....
2020-07-28 12:27:58  Epoch 0...
        train data: loss:107503.6215, mse: 2.0768;
        evaluation reslut: mse: 1.2466; rmse: 1.1165; mae: 0.9691;
model save
******************************
2020-07-28 12:28:13  Epoch 1...
        train data: loss:80552.2573, mse: 1.5561;
        evaluation reslut: mse: 1.0296; rmse: 1.0147; mae: 0.8384;
model save
******************************
2020-07-28 12:28:29  Epoch 2...
        train data: loss:70202.6199, mse: 1.3562;
        evaluation reslut: mse: 0.9926; rmse: 0.9963; mae: 0.8146;
model save
******************************

Framework Design

An overview of the package dir:

Data Preprocessing

After data processing, one record of the training/validation/test dataset is:

user_id, item_id, ratings

For example the training data triples are stored as Train.npy, Train_Score.npy in dataset/Digital_Music_data/train/.

The review information of users and items are preprocessed in the following format:

user_id
user_doc: the word index sequence of the document of the user, [w1, w2, ... wn]
user_reviews list: the list of all the review of the user, [[w1,w2..], [w1,w2,..],...[w1,w2..]]
user_item2id: the item ids that the user have purchased, [id1, id2,...]

The same as the items. Hence in the code, we orgnize our batch data as:

uids, iids, user_reviews, item_reviews, user_item2id, item_user2id, user_doc, item_doc

This is all the information involved in review-based recommendation, researchers can utilize this data format to build own models. Note that the review in validation/test dataset is excluded.

Note that the review processing methods are usually different among these papers (e.g., the vocab, padding), which would influence their performance. In this repo, to be fair, we adopt the same pre-poressing approach for all the methods. Hence the performance may be not consistent with the origin papers.

Model Details

In order to make our framework more extensible, we define three modules in our framework:

User/Item Representation Learning Layer (in models/*py): the main part of most baseline methods, such as the CNN encoder in DeepCoNN.
Fusion Layer in framework/fusion.py: combine the user/item different features (e.g., ID feature and review/doc feature), and then fuse the user and item feature into one feature vector, we pre-define the following methods:
- sum
- add
- concatenation
- self attention
Prediction Layer in framework/prediction.py: prediction the score that user towards item (i.e., a regression layer), we pre-define the following rating prediction layers:
- (Neural) Factorization Machine
- Latent Factor Model
- MLP

Hence, researchers could build their models in user/item representation learning layer.

Note that if you would like to add new method or datset, please remember to declare in the __init__.py

Citation

If you use the code, please cite:

@inproceedings{liu2019nrpa,
  title={NRPA: Neural Recommendation with Personalized Attention},
  author={Liu, Hongtao and Wu, Fangzhao and Wang, Wenjun and Wang, Xianchen and Jiao, Pengfei and Wu, Chuhan and Xie, Xing},
  booktitle={Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval},
  pages={1233--1236},
  year={2019}
}

huangtianan/Neu-Review-Rec