TranSearch

A pytorch and tensorflow GPU implementation for our paper:

Yangyang Guo, Zhiyong Cheng, Liqiang Nie, Xin-Shun Xu, Mohan Kankanhalli (2018). Multi-modal Preference Modeling for Product Search. In Proceedings of ACM MM 2018.

Please cite our MM'18 paper if you use our codes. Thanks!

You can download the Amazon Dataset from http://jmcauley.ucsd.edu/data/amazon.

The requirements are as follows:

* python==3.6
* pandas==0.24.2
* numpy==1.16.2
* pytorch==0.4.0
* tensorflow==1.7
* gensim==3.7.1
* tensorboardX==1.6

Example to run:

• Make sure the review data, meta data and image features are in the same directory. Be careful with the review, meta and image file name.

• Preprocessing data. Noted that the category >Clothing contains both the >Men's Clothing and >Women's Clothing.

   python scripts/preprocess.py  --count=5 --is_clothing
  

• We leverage the PV-DM model to convert queries and product representations to the same latent space. Besides, we extract the image features here. Be sure to use the correct image feature weights.

   python script/doc2vec_img2dict.py 
  

• Pre-training is optional to train the final model.

   python pytorch/pre_train.py --embed_size=128 --neg_number=20
  

• Train and test the framework.

   python pytorch/TranSearch.py --mode='text' --embed_size=128