FudanCISL / FIRE

FIRE forked from official implementation

FIRE

This repository is the implementation of the paper: FIRE: Fast Incremental Recommendation with Graph Signal Processing (TheWebConf 2022).

If you use the code in your work, please cite our paper.

How to run the code

Step 1: Check the compatibility of your python packages, we recommend you to use the following setting that has passed our test if you want to reproduce the results.

numpy == 1.19.4
pandas == 1.1.4
python == 3.6.11
scipy == 1.5.3
sklearn == 0.23.2
sparsesvd == 0.2.2

Step 2: prepare the dataset.

• If you use the default datasets (Movielens 1M or Douban Movie), please unzip the dataset under the directory dataset/ first;

• If you use your own dataset, you should make sure that the format of your dataset is compatible with the setting (i.e. the following variables) in dataloader.py

  • sep : the separation symbol between columns in dataset (default: '\t')
  • header_name: the header name of data frame generated from dataset (default: ['u', 'i', 'r', 't'])
  • pos_type: The type of positive interactions (default: [4.0, 5.0]).

  Note: You should disable the comment of Line 23-27 in dataloader.py to generate a new column named 'm' when loading data if you use your own dataset.

Step 3: run the model.

• For Movielens 1M dataset:

  python main.py --dataset ml1m  --use_user_si --use_item_si

• For Douban Movie dataset:

  python main.py --dataset douban_movie --num_his_month 9 --num_cur_month 1 --decay_factor 1e-8 --pri_factor 128 --alphas '[0.2,1.0,0,0.2]' --use_item_si --user_threshold 0 --item_threshold 0.6