Our team ranks the second place on the final leaderboard with an excellent performance very close to the first place. The relevant solution report can be found in here.
The requirements of experiment environment are listed in requirements.txt.
We totally follow the structure of xmrec's sample code for our own model design. Thus it is very easy to reproduce for version of lightgcn.
Before running train script, you should perform preprocessing by running script merge_train.py. After that, train script is easy to run by referring to the samples below.
train_baseline.py is the script for training our model that is taking one target market.
Here is a sample train script from training model using data of t1 without word2vec:
python train_baseline.py --tgt_market t1 --src_markets none --tgt_market_valid DATA/t1/valid_run.tsv --tgt_market_test DATA/t1/test_run.tsv --exp_name lightgcn_without --num_epoch 60 --cuda --model_name LightGCN_no_w2v
Here is a sample train script from training model using data of t2 without word2vec:
python train_baseline.py --tgt_market t2 --src_markets none --tgt_market_valid DATA/t2/valid_run.tsv --tgt_market_test DATA/t2/test_run.tsv --exp_name lightgcn_without --num_epoch 50 --cuda --tgt_market_valid_pos DATA/t2/valid_qrel.tsv --model_name LightGCN_no_w2v
Here is a sample train script from training model using data of t1 with word2vec:
python train_baseline.py --tgt_market t1 --src_markets none --tgt_market_valid DATA/t1/valid_run.tsv --tgt_market_test DATA/t1/test_run.tsv --exp_name lightgcn_with --num_epoch 60 --cuda --model_name LightGCN_with_w2v
Here is a sample train script from training model using data of t2 with word2vec:
python train_baseline.py --tgt_market t2 --src_markets none --tgt_market_valid DATA/t2/valid_run.tsv --tgt_market_test DATA/t2/test_run.tsv --exp_name lightgcn_with --num_epoch 50 --cuda --tgt_market_valid_pos DATA/t2/valid_qrel.tsv --model_name LightGCN_with_w2v
These four samples can be executed seperately to accelerate training procedure.
After training your model, the scripts prints the directories of model and index checkpoints as well as the run files for the validation and test data as below.
Run output files:
--validation: ./merge/t1/LightGCN_with_w2v_valid.tsv
--test: ./merge/t1/LightGCN_with_w2v_test.tsv
Experiment finished successfully!
After getting corresponding result files, result files will be input into other tree-like models as ranking features.
You can run scripts below:
#6.recall feature
python recall_feature.py
python recall_feature_2.py
# Two python scripts above can run in parellel
#7.lightgbm Classifier model
python class_lgb_model.py
#8.xgboost Classifier model
python class_xgb_model.py
#9.catboost Classifier model
python class_cat_model.py
#10.lightgbm Rank model
python rank_lgb_model.py
# Four python scripts(#7,#8,#9,#10) above can run in parellel
Finally, by running trans_file.py to aggregate multiple results, you can get final submissions we have submitted in path ./submission.