{
  "id": 382910,
  "title": "168th place solution(Bronze medal)",
  "url": "/competitions/otto-recommender-system/discussion/382910",
  "author_name": "",
  "post_date": "2023-02-01T14:03:03.913660800Z",
  "votes": 9,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I feel OTTO is most difficult competition I have ever participated.<br>\nUntil two days before the end of the competition, heuristics was unable to beat the ranker model, but by creating features for the covis matrix and increasing the number of covis matrix candidates to the maximum (top50), it was finally able to beat the ranker model.</p>\n<p>Thank you for my team, <a href=\"https://www.kaggle.com/coffeemountain\" target=\"_blank\">@coffeemountain</a> , <a href=\"https://www.kaggle.com/hiroki1018\" target=\"_blank\">@hiroki1018</a>  without them, I would not have won the bronze medal.</p>\n<h4>Environment</h4>\n<ul>\n<li>google colab pro+<br>\nWe used GPU premium 80GB only for large memory even thought we didn’t use GPU. <br>\nSo we charged a lot for google colab😂</li>\n</ul>\n<h4>Solustion</h4>\n<h4>stage1</h4>\n<h5>- candidate: 50 (chris baseline)</h5>\n<h5>- features</h5>\n<h6>　user features : 13</h6>\n<ul>\n<li>'n_items_by_session_ordered',</li>\n<li>'average_time_between_clicks',</li>\n<li>'n_items_by_session_carted',</li>\n<li>'user_user_count',</li>\n<li>'day_of_week_made_last_activity',</li>\n<li>'series_time',</li>\n<li>'n_items_by_session_clicked',</li>\n<li>'average_hour_carts',</li>\n<li>'n_real_sessions_within_1800s',</li>\n<li>'average_hour_clicks',</li>\n<li>'user_item_count',</li>\n<li>'average_n_items_by_real_sessions_within_1800s',</li>\n<li>'user_buy_ratio',</li>\n</ul>\n<h6>　item feature : 24</h6>\n<ul>\n<li>'aid_train_ordered_times',</li>\n<li>'aid_init_clicked_train_date',</li>\n<li>'item_item_count',</li>\n<li>'aid_train_clicked_times',</li>\n<li>'item_user_count',</li>\n<li>'average_hour_item_carted',</li>\n<li>'aid_init_carted_train_date',</li>\n<li>'aid_train_carted_times',</li>\n<li>'before_1day_ordered_ratio',</li>\n<li>'before_1day_carted_ratio',</li>\n<li>'before_1day_clicked_ratio',</li>\n<li>'before_1day_clicked_flag',</li>\n<li>'before_1day_carted_flag',</li>\n<li>'before_1day_ordered_flag',</li>\n<li>'item_ever_been_bought_train',</li>\n<li>'aid_multi_click_user_ratio',</li>\n<li>'average_hour_item_ordered',</li>\n<li>'weekday0_clicked_popular_ratio',</li>\n<li>'aid_init_ordered_train_date',</li>\n<li>'item_buy_ratio',</li>\n<li>'popularity_ranking_in_train',</li>\n<li>'before_5day_carted_ratio',</li>\n<li>'before_5day_clicked_ratio',</li>\n<li>'before_5day_ordered_ratio',</li>\n</ul>\n<h6>　interaction feature: 16</h6>\n<ul>\n<li>'has_been_carted_flag',</li>\n<li>'items_count_by_session_action',</li>\n<li>'session_first_aid',</li>\n<li>'session_last_aid',</li>\n<li>'series_aid_flag',</li>\n<li>'more_than_24hours_time_gap_flag',</li>\n<li>'click_covis_matrix_weight',</li>\n<li>'aid_in_session_flag',</li>\n<li>'click_covis_matrix_top50_ranking',</li>\n<li>'reverse_test_aid_number',</li>\n<li>'reverse_test_aid_ratio',</li>\n<li>'action_interval_in_session',</li>\n<li>'reverse_action_interval_accumulation'</li>\n</ul>\n<h4>stage2</h4>\n<p>Ranking: lightgbm</p>\n<ul>\n<li>'objective': 'lambdarank',</li>\n<li>'metric': 'ndcg', </li>\n<li>'ndcg_eval_at': 20,</li>\n<li>'n_estimators': 200,</li>\n<li>'boosting_type': 'dart',</li>\n</ul>\n<h4>Not worked</h4>\n<ul>\n<li>Future features <br>\nex) count  actions in future Ndays </li>\n<li>optuna<br>\netc.</li>\n</ul>",
  "messages": [
    {
      "id": "2125180",
      "postDate": "02/01/2023 14:03:03",
      "content": "<p>I feel OTTO is most difficult competition I have ever participated.<br>\nUntil two days before the end of the competition, heuristics was unable to beat the ranker model, but by creating features for the covis matrix and increasing the number of covis matrix candidates to the maximum (top50), it was finally able to beat the ranker model.</p>\n<p>Thank you for my team, <a href=\"https://www.kaggle.com/coffeemountain\" target=\"_blank\">@coffeemountain</a> , <a href=\"https://www.kaggle.com/hiroki1018\" target=\"_blank\">@hiroki1018</a>  without them, I would not have won the bronze medal.</p>\n<h4>Environment</h4>\n<ul>\n<li>google colab pro+<br>\nWe used GPU premium 80GB only for large memory even thought we didn’t use GPU. <br>\nSo we charged a lot for google colab😂</li>\n</ul>\n<h4>Solustion</h4>\n<h4>stage1</h4>\n<h5>- candidate: 50 (chris baseline)</h5>\n<h5>- features</h5>\n<h6>　user features : 13</h6>\n<ul>\n<li>'n_items_by_session_ordered',</li>\n<li>'average_time_between_clicks',</li>\n<li>'n_items_by_session_carted',</li>\n<li>'user_user_count',</li>\n<li>'day_of_week_made_last_activity',</li>\n<li>'series_time',</li>\n<li>'n_items_by_session_clicked',</li>\n<li>'average_hour_carts',</li>\n<li>'n_real_sessions_within_1800s',</li>\n<li>'average_hour_clicks',</li>\n<li>'user_item_count',</li>\n<li>'average_n_items_by_real_sessions_within_1800s',</li>\n<li>'user_buy_ratio',</li>\n</ul>\n<h6>　item feature : 24</h6>\n<ul>\n<li>'aid_train_ordered_times',</li>\n<li>'aid_init_clicked_train_date',</li>\n<li>'item_item_count',</li>\n<li>'aid_train_clicked_times',</li>\n<li>'item_user_count',</li>\n<li>'average_hour_item_carted',</li>\n<li>'aid_init_carted_train_date',</li>\n<li>'aid_train_carted_times',</li>\n<li>'before_1day_ordered_ratio',</li>\n<li>'before_1day_carted_ratio',</li>\n<li>'before_1day_clicked_ratio',</li>\n<li>'before_1day_clicked_flag',</li>\n<li>'before_1day_carted_flag',</li>\n<li>'before_1day_ordered_flag',</li>\n<li>'item_ever_been_bought_train',</li>\n<li>'aid_multi_click_user_ratio',</li>\n<li>'average_hour_item_ordered',</li>\n<li>'weekday0_clicked_popular_ratio',</li>\n<li>'aid_init_ordered_train_date',</li>\n<li>'item_buy_ratio',</li>\n<li>'popularity_ranking_in_train',</li>\n<li>'before_5day_carted_ratio',</li>\n<li>'before_5day_clicked_ratio',</li>\n<li>'before_5day_ordered_ratio',</li>\n</ul>\n<h6>　interaction feature: 16</h6>\n<ul>\n<li>'has_been_carted_flag',</li>\n<li>'items_count_by_session_action',</li>\n<li>'session_first_aid',</li>\n<li>'session_last_aid',</li>\n<li>'series_aid_flag',</li>\n<li>'more_than_24hours_time_gap_flag',</li>\n<li>'click_covis_matrix_weight',</li>\n<li>'aid_in_session_flag',</li>\n<li>'click_covis_matrix_top50_ranking',</li>\n<li>'reverse_test_aid_number',</li>\n<li>'reverse_test_aid_ratio',</li>\n<li>'action_interval_in_session',</li>\n<li>'reverse_action_interval_accumulation'</li>\n</ul>\n<h4>stage2</h4>\n<p>Ranking: lightgbm</p>\n<ul>\n<li>'objective': 'lambdarank',</li>\n<li>'metric': 'ndcg', </li>\n<li>'ndcg_eval_at': 20,</li>\n<li>'n_estimators': 200,</li>\n<li>'boosting_type': 'dart',</li>\n</ul>\n<h4>Not worked</h4>\n<ul>\n<li>Future features <br>\nex) count  actions in future Ndays </li>\n<li>optuna<br>\netc.</li>\n</ul>",
      "rawMarkdown": "I feel OTTO is most difficult competition I have ever participated.\nUntil two days before the end of the competition, heuristics was unable to beat the ranker model, but by creating features for the covis matrix and increasing the number of covis matrix candidates to the maximum (top50), it was finally able to beat the ranker model.\n\nThank you for my team, @coffeemountain , @hiroki1018  without them, I would not have won the bronze medal.\n\n\n\n\n#### Environment\n* google colab pro+\nWe used GPU premium 80GB only for large memory even thought we didn’t use GPU. \nSo we charged a lot for google colab😂\n\n#### Solustion\n\n#### stage1\n##### - candidate: 50 (chris baseline)\n##### - features\n######　user features : 13\n- 'n_items_by_session_ordered',\n- 'average_time_between_clicks',\n- 'n_items_by_session_carted',\n- 'user_user_count',\n- 'day_of_week_made_last_activity',\n- 'series_time',\n- 'n_items_by_session_clicked',\n- 'average_hour_carts',\n- 'n_real_sessions_within_1800s',\n- 'average_hour_clicks',\n- 'user_item_count',\n- 'average_n_items_by_real_sessions_within_1800s',\n- 'user_buy_ratio',\n\n\n######　item feature : 24\n- 'aid_train_ordered_times',\n- 'aid_init_clicked_train_date',\n- 'item_item_count',\n- 'aid_train_clicked_times',\n- 'item_user_count',\n- 'average_hour_item_carted',\n- 'aid_init_carted_train_date',\n- 'aid_train_carted_times',\n- 'before_1day_ordered_ratio',\n- 'before_1day_carted_ratio',\n- 'before_1day_clicked_ratio',\n- 'before_1day_clicked_flag',\n- 'before_1day_carted_flag',\n- 'before_1day_ordered_flag',\n- 'item_ever_been_bought_train',\n- 'aid_multi_click_user_ratio',\n- 'average_hour_item_ordered',\n- 'weekday0_clicked_popular_ratio',\n- 'aid_init_ordered_train_date',\n- 'item_buy_ratio',\n- 'popularity_ranking_in_train',\n- 'before_5day_carted_ratio',\n- 'before_5day_clicked_ratio',\n- 'before_5day_ordered_ratio',\n\n######　interaction feature: 16\n- 'has_been_carted_flag',\n- 'items_count_by_session_action',\n- 'session_first_aid',\n- 'session_last_aid',\n- 'series_aid_flag',\n- 'more_than_24hours_time_gap_flag',\n- 'click_covis_matrix_weight',\n- 'aid_in_session_flag',\n- 'click_covis_matrix_top50_ranking',\n- 'reverse_test_aid_number',\n- 'reverse_test_aid_ratio',\n- 'action_interval_in_session',\n- 'reverse_action_interval_accumulation'\n\n#### stage2\nRanking: lightgbm\n- 'objective': 'lambdarank',\n- 'metric': 'ndcg', \n- 'ndcg_eval_at': 20,\n- 'n_estimators': 200,\n- 'boosting_type': 'dart',\n\n####  Not worked\n- Future features \n ex) count  actions in future Ndays \n- optuna\netc.",
      "votes": null
    },
    {
      "id": "2125441",
      "postDate": "02/01/2023 16:57:25",
      "content": "<p>Congratulate! are u trined the model in kaggle notebook? i feel 30GB ram is not enough for such feature size</p>",
      "rawMarkdown": "Congratulate! are u trined the model in kaggle notebook? i feel 30GB ram is not enough for such feature size",
      "votes": null
    },
    {
      "id": "2126161",
      "postDate": "02/02/2023 05:47:06",
      "content": "<p>Great work on this and thank you for the write up!<br>\nI'm curious, what features exactly did end up yielding the desired result in the end?<br>\nThe Devastator.</p>",
      "rawMarkdown": "Great work on this and thank you for the write up!\nI'm curious, what features exactly did end up yielding the desired result in the end?\nThe Devastator.",
      "votes": null
    },
    {
      "id": "2126778",
      "postDate": "02/02/2023 13:01:10",
      "content": "<p>Thank you for your comment!</p>\n<p>In our case, covistation matrix boost LB score, because there are many covistation candidate in our candidate.</p>",
      "rawMarkdown": "Thank you for your comment!\n\nIn our case, covistation matrix boost LB score, because there are many covistation candidate in our candidate.",
      "votes": null
    },
    {
      "id": "2126799",
      "postDate": "02/02/2023 13:06:10",
      "content": "<p>Thank you for your comment!<br>\nI didn't try Kaggle notebook, I think it happen out of memory.</p>",
      "rawMarkdown": "Thank you for your comment!\nI didn't try Kaggle notebook, I think it happen out of memory.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2125441,
      "author_name": "xianzwaikato",
      "author_url": "",
      "post_date": "02/01/2023 16:57:25",
      "content": "<p>Congratulate! are u trined the model in kaggle notebook? i feel 30GB ram is not enough for such feature size</p>",
      "votes": null,
      "replies": [
        {
          "id": 2126799,
          "author_name": "mujrush",
          "author_url": "",
          "post_date": "02/02/2023 13:06:10",
          "content": "<p>Thank you for your comment!<br>\nI didn't try Kaggle notebook, I think it happen out of memory.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2126161,
      "author_name": "thedevastator",
      "author_url": "",
      "post_date": "02/02/2023 05:47:06",
      "content": "<p>Great work on this and thank you for the write up!<br>\nI'm curious, what features exactly did end up yielding the desired result in the end?<br>\nThe Devastator.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2126778,
          "author_name": "mujrush",
          "author_url": "",
          "post_date": "02/02/2023 13:01:10",
          "content": "<p>Thank you for your comment!</p>\n<p>In our case, covistation matrix boost LB score, because there are many covistation candidate in our candidate.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2125180": "I feel OTTO is most difficult competition I have ever participated.\nUntil two days before the end of the competition, heuristics was unable to beat the ranker model, but by creating features for the covis matrix and increasing the number of covis matrix candidates to the maximum (top50), it was finally able to beat the ranker model.\n\nThank you for my team, @coffeemountain , @hiroki1018  without them, I would not have won the bronze medal.\n\n\n\n\n#### Environment\n* google colab pro+\nWe used GPU premium 80GB only for large memory even thought we didn’t use GPU. \nSo we charged a lot for google colab😂\n\n#### Solustion\n\n#### stage1\n##### - candidate: 50 (chris baseline)\n##### - features\n######　user features : 13\n- 'n_items_by_session_ordered',\n- 'average_time_between_clicks',\n- 'n_items_by_session_carted',\n- 'user_user_count',\n- 'day_of_week_made_last_activity',\n- 'series_time',\n- 'n_items_by_session_clicked',\n- 'average_hour_carts',\n- 'n_real_sessions_within_1800s',\n- 'average_hour_clicks',\n- 'user_item_count',\n- 'average_n_items_by_real_sessions_within_1800s',\n- 'user_buy_ratio',\n\n\n######　item feature : 24\n- 'aid_train_ordered_times',\n- 'aid_init_clicked_train_date',\n- 'item_item_count',\n- 'aid_train_clicked_times',\n- 'item_user_count',\n- 'average_hour_item_carted',\n- 'aid_init_carted_train_date',\n- 'aid_train_carted_times',\n- 'before_1day_ordered_ratio',\n- 'before_1day_carted_ratio',\n- 'before_1day_clicked_ratio',\n- 'before_1day_clicked_flag',\n- 'before_1day_carted_flag',\n- 'before_1day_ordered_flag',\n- 'item_ever_been_bought_train',\n- 'aid_multi_click_user_ratio',\n- 'average_hour_item_ordered',\n- 'weekday0_clicked_popular_ratio',\n- 'aid_init_ordered_train_date',\n- 'item_buy_ratio',\n- 'popularity_ranking_in_train',\n- 'before_5day_carted_ratio',\n- 'before_5day_clicked_ratio',\n- 'before_5day_ordered_ratio',\n\n######　interaction feature: 16\n- 'has_been_carted_flag',\n- 'items_count_by_session_action',\n- 'session_first_aid',\n- 'session_last_aid',\n- 'series_aid_flag',\n- 'more_than_24hours_time_gap_flag',\n- 'click_covis_matrix_weight',\n- 'aid_in_session_flag',\n- 'click_covis_matrix_top50_ranking',\n- 'reverse_test_aid_number',\n- 'reverse_test_aid_ratio',\n- 'action_interval_in_session',\n- 'reverse_action_interval_accumulation'\n\n#### stage2\nRanking: lightgbm\n- 'objective': 'lambdarank',\n- 'metric': 'ndcg', \n- 'ndcg_eval_at': 20,\n- 'n_estimators': 200,\n- 'boosting_type': 'dart',\n\n####  Not worked\n- Future features \n ex) count  actions in future Ndays \n- optuna\netc.",
    "2125441": "Congratulate! are u trined the model in kaggle notebook? i feel 30GB ram is not enough for such feature size",
    "2126161": "Great work on this and thank you for the write up!\nI'm curious, what features exactly did end up yielding the desired result in the end?\nThe Devastator.",
    "2126778": "Thank you for your comment!\n\nIn our case, covistation matrix boost LB score, because there are many covistation candidate in our candidate.",
    "2126799": "Thank you for your comment!\nI didn't try Kaggle notebook, I think it happen out of memory."
  },
  "source": "meta"
}