/Iterative-Relabeled-One-Class-Collaborative-Filtering-against-Noisy-Interactions

Primary LanguagePython

Iterative-Relabeled-One-Class-Collaborative-Filtering-against-Noisy-Interactions

This is a Pytorch implementation of the model described in our paper:

Z. Wang, Q. Xu, Z. Yang, X. Cao and Q. Huang. Implicit Feedbacks are Not Always Favorable: Iterative Relabeled One-Class Collaborative Filtering against Noisy Interactions. MM2021.

Dependencies

  • Pytorch >= 1.5.1
  • numpy

Data

We convert the datasets ML100K, ML1M and Netflix to our train and test files in the data/ folder.

To generate implicit training data, we randomly select $n_r$ ratings for each user as observed interactions, no matter whether the user likes the item or not. Meanwhile, we binarize the rest ratings according to a threshold of 4 to evaluate the performance in predicting true user preference, which is unavailable in implicit datasets. Note that users with less than $n_r$ ratings and items not appearing in the training set are filtered out.

The ratings are stored in the files *.lsvm. The data format of the line user_id is item_id:rating. The numeric ratings range from 1 to 5.

Train

Here is an example to generate the new data and train the model.

python main.py

Citation

Please cite our paper if you use this code in your own work:

@inproceedings{DBLP:conf/mm/WangX0CH21,
  author    = {Zitai Wang and
               Qianqian Xu and
               Zhiyong Yang and
               Xiaochun Cao and
               Qingming Huang},
  title     = {Implicit Feedbacks are Not Always Favorable: Iterative Relabeled One-Class Collaborative Filtering against Noisy Interactions},
  booktitle = {{ACM} Multimedia Conference},
  pages     = {3070--3078},
  year      = {2021},
}