Multi-Task Pairwise Ranking (MTPR) aims to address the discrepancy during the training process of cold-start recommendation. The discrepancy results in suboptimal item representation for multimedia recommendation.
We devised dual item representations consisting of two vectors, where N-rep indicates the normal representation that corresponds to all item representations and C-rep indicates the counterfactual representation that assumes the item is cold. We then proposed multi-task pairwise ranking, a loss function composed of four pairwise ranking branches, to optimize the representations with the target of modeling the correlations between non-cold and cold items.
This our official implementation for the paper:
Xiaoyu Du, Xiang Wang, Xiangnan He, Zechao Li, Jinhui Tang, and Tat-Seng Chua, How to Learn Item Representation for Cold-Start Multimedia Recommendation?, in Proceedings of MM'20.
If you use the codes, please cite our paper. Thanks!
- Python==3.5
- Pytorch==1.4
- numpy, scipy, argparse, logging, sklearn, tqdm
python main.python
We provide the dataset "amazon" in amazon/men.npy
.
>>> import numpy as np
>>> data = np.load('amazon/men.npy', allow_pickle=True)
>>> data.item().keys()
dict_keys(['train', 'feat', 'val', 'test'])
>>> len(data.item()['train'])
27044
>>> len(data.item()['val'])
27044
>>> len(data.item()['test'])
27044
>>> data.item()['feat'].shape
(86506, 64)