Skip to content

Latest commit

 

History

History
119 lines (91 loc) · 4.91 KB

README.md

File metadata and controls

119 lines (91 loc) · 4.91 KB
Raw

Solution of Kakao Recoteam to Recsys Challenge 2022

Refer to https://www.recsyschallenge.com/2022/

Experimental Environments:

  • We tested on Ubuntu 18.04
  • We tested on Python==3.9.1, torch==1.11.0 with cuda 11.3
  • GCC>=7.2.1 is required since BERT4REC dataset builder uses C++-17

Before Running

  • Making Dataset (Download link): put unzipped data(candidate_items.csv, train_purchase.csv, ...,) in the directory ./data
  • Install Dependencies:
    pip3 install -r requirements.txt
pip3 install torch-scatter -f https://data.pyg.org/whl/torch-1.11.0+cu113.html
pip3 install torch-sparse -f https://data.pyg.org/whl/torch-1.11.0+cu113.html
pip3 install torch-geometric

We have several models: MLP, BERT4Rec, GRU, GNN, GRU + GNN (grun), and PCos.

Preprocessing

  • For submission (leaderboard, final)
python preprocessing.py --submit=True

We used sessions/purchases in last month in train data as validation/test for internal validation. If we set --submit=False, we build preprocessed for internal validation.

python preprocessing.py --submit=False

Model Training

Train each model and save model weights. ($bash ./run.sh)

Note that the save_fname rule is as follows. {model_name}-{extra_kwargs_1}-{extra_kwargs_2}-{extra_kwargs_k}.pt

#!/bin/bash
mkdir save
for i in {1..20}
do
    python module/models/gru.py --submit=True --num_workers=4 --save_fname=save/gru.pt --seed=$i
    python module/models/gnn.py --submit=True --num_workers=4 --save_fname=save/gnn.pt --seed=$i
    python module/models/grun.py --submit=True --num_workers=4 --save_fname=save/grun.pt --seed=$i
    python module/models/grun.py --submit=True --num_workers=4 --all=True --save_fname=save/grun-all.pt --seed=$i
done

MLP, and BERT4rec have different implementation so we includede it in separate directories. MLP in directory (./mlp) and (./BERT4REC). To train/run those models, check MLP README (mlp/README.md) and BERT4rec readme(./BERT4Rec/README.md).

BERT4REC has dependencies on our internal workflow so we removed the dependencies and included it in a separate directory(./BERT4REC). See BERT4REC/README.md.

Unnormalized predicted scores (logits) for those models are stored in ./logits/{val, test, leader, final}/{mlp, bert4rec}_timestamp.logit and have same format with other models. Model ensemble below can be run without MLP/BERT4REC yet we achieved highest scores with those two models included.

Ensemble Models

Saved models are as follows.

In [1]: from glob import glob
In [2]: saves = glob('save/*.pt*')
In [3]: saves
Out[3]:
['save/gru.pt.1656302353',
 'save/gnn.pt.1656302415',
 'save/grun.pt.1656302622',
 'save/grun-all.pt.1656304624'

Generate logits and ensemble all of them to generate recommended lists for leaderboard sessions.

Please note that predictions are made from parsed logit files, so training process should be done beforehand. Note that MLP/BERT4REC themselves generate logits.

python module/models/ensemble.py --fork_fallback=True --submit=True --kind=leader  # recommendations for leaderboard sessions
python module/models/ensemble.py --fork_fallback=True --submit=True --kind=final  # recommendations for final sessions

submit option determines the data we use to make ensemble.

  • If --submit=False is set to true, it generates logits for internal validation. Otherwise, it generates logits for submission.

Note: In final submission, we employ weighted sum when model ensemble. check ./notebook/run.py for our emsemble logic

kind option determines the target for recommendation results: leader and final. (val and te for internal validation)

  • If --kind=leader, it generates recommendation results on leaderboard sessions.
  • If --kind=final, it generates recommendation results on final sessions.
  • If kind is not set, it will default to val.

Contributors

License

This software is licensed under the Apache 2 license, quoted below.

Copyright 2022 Kakao Corp. https://www.kakaocorp.com

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this project except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0.

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.