PSSL

Source code of CIKM2021 Long Paper "PSSL: Self-supervised Learning for Personalized Search with Contrastive Sampling". It consists of the pre-training stage ./pretrain and the fine-tuning stage ./finetune.

Requirements

I test the code with the following packages. Other versions may also work, but I'm not sure.

• Python 3.6
  
• Pytorch 1.3.1 (with GPU support)

Usage

pre-training

At the pre-training stage, we first extract data pairs from query logs, as shown in /datasample/***_pair.txt. Then, we train the parameters with the contrastive learning framework by:

python pretrain/prepare_load.py

fine-tuning

We initial the encoders with pre-trained parameters and train the ranking model by:

python finetune/dataset_new.py

Citations

If you use the code, please cite the following paper:

@inproceedings{ZhouDYW21,
  author    = {Yujia Zhou and
               Zhicheng Dou and
               Yutao Zhu and
               Ji{-}Rong Wen},
  title     = {PSSL: Self-supervised Learning for Personalized Search with Contrastive Sampling},
  booktitle = {{CIKM}},
  publisher = {{ACM}},
  year      = {2021}
}