Recsys Challenge

Requirements

• python3
• torch>=1.1.0
• pytorch-pretrained-bert==0.6.2
• tensorboardx==2.0
• tqdm==4.44

Usage

Prepare

python data_count.py

This program will count the language, hashtag and user information. The path of the source data is assigned in the code.

Train

python train.py \
--device 0 \
--batch 2000 \
--model autoint \
--model_name exp1 \
--label retweet \
--data_name all \
--epoch 6 \
--lr 1e-3 \
--weight_decay 1e-6 \
--dropout 0.5 \
--n_workers 0 \
--use_user_info \
--save_latest

If using [--use_user_info], the model will use user's information collected from training and validation set, which will consume huge RAM memory, and only when the test set user set coincides a lot with training set can this option improve the model. n_workers can be set higher to speed up data processing by multiprocessing. Note that the memory occupation is proportional to the number of processes.

The path of the training set and validation set is assigned in config.py.

Prediction

python test.py --device 0 model_name exp1 --label retweet --data_name all --use_user_info --make_prediction

This is used for submission. Eliminate --make_prediction for evaluation on the validation dataset. [--use_user_info] should keep consistent with training. In practice, we split the test data into two parts, one where all user appear at least once in training data and vice versa, and apply corresponding models.

The path of the test set is assigned in config.py.