![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6434827%2F819ad8ee7a0a74803a927f6cd0ffb180%2F2023-02-04%2012.10.28.png?generation=1675480262988727&alt=media\")
First of all, I would like to thank kaggle and OTTO staff for organizing great competition. And I would also like to thank my teammates.（@zakopur0, @irrohas, @t88take, @chiakiichimura）
\n
Google Colab Pro+
", "messages": [ { "id": 2128759, "postDate": "2023-02-04T03:11:14.373Z", "content": "First of all, I would like to thank kaggle and OTTO staff for organizing great competition. And I would also like to thank my teammates.（@zakopur0, @irrohas, @t88take, @chiakiichimura）
\n
Google Colab Pro+
", "rawMarkdown": "First of all, I would like to thank kaggle and OTTO staff for organizing great competition. And I would also like to thank my teammates.（@zakopur0, @irrohas, @t88take, @chiakiichimura）\n\n#### Overview \n- 2-Stage Recommendation Model\n**Stage_1: Candidate generation**\n- Generate candidates from various models.（Candidates are selected by arranging positive rate.)\n- Create several Sequential Recommendation models using Recbole.\n**Stage_2: Ranking Model**\n- Stacking the output of two models, \"Re-Action aid model\" and \"Related aid model\".\n- 4w labels used for training, CV verified by GroupKFold(session)\n- About 100 candidates per session. About 100 features.\n**Ensemble**\n- Ensemble of 6 different models with different TOP:N of candidate selection.\n- Ensemble with team member submissions.\n\n\n\n#### What I worked on but didn't work\n- Post-processing sold-out aids, big-sell aids.(485256, 33343, 660655, etc.）\n- Candidate generation by graph NN model.（LightGCN）\n- Ensemble of different GBDT models.（XGBoost Classifier, LightGBM Classifier, LightGBM Ranker)\n\n\n#### Environment\nGoogle Colab Pro+\n", "votes": 15 } ], "comments": [], "raw_markdown_by_id": { "2128759": "First of all, I would like to thank kaggle and OTTO staff for organizing great competition. And I would also like to thank my teammates.（@zakopur0, @irrohas, @t88take, @chiakiichimura）\n\n#### Overview \n- 2-Stage Recommendation Model\n**Stage_1: Candidate generation**\n- Generate candidates from various models.（Candidates are selected by arranging positive rate.)\n- Create several Sequential Recommendation models using Recbole.\n**Stage_2: Ranking Model**\n- Stacking the output of two models, \"Re-Action aid model\" and \"Related aid model\".\n- 4w labels used for training, CV verified by GroupKFold(session)\n- About 100 candidates per session. About 100 features.\n**Ensemble**\n- Ensemble of 6 different models with different TOP:N of candidate selection.\n- Ensemble with team member submissions.\n\n\n\n#### What I worked on but didn't work\n- Post-processing sold-out aids, big-sell aids.(485256, 33343, 660655, etc.）\n- Candidate generation by graph NN model.（LightGCN）\n- Ensemble of different GBDT models.（XGBoost Classifier, LightGBM Classifier, LightGBM Ranker)\n\n\n#### Environment\nGoogle Colab Pro+\n" } }