Thanks OTTO & Kaggle to host such interesting competition, my solution is straight forward, just candidate generation + reranker, a short summary as below.
\nFinally, 3 models(clicks/carts/orders) is trained by LightGBM(lambdarank)
", "messages": [ { "id": "2124826", "postDate": "02/01/2023 08:32:34", "content": "Thanks OTTO & Kaggle to host such interesting competition, my solution is straight forward, just candidate generation + reranker, a short summary as below.
\nFinally, 3 models(clicks/carts/orders) is trained by LightGBM(lambdarank)
", "rawMarkdown": "Thanks OTTO & Kaggle to host such interesting competition, my solution is straight forward, just candidate generation + reranker, a short summary as below.\n\n###### Candidate Generation\n\n- **re-visit** - all visited items\n- **co-visit1** - based on the public notebook with 2 improvement. Thanks @radek1 and @cdeotte \n 1. candidate generation by weight decay, e.g. a session has N items(idx1 by the reverse ts) & each item has M pairs of co-visit item(idx2 by time weight or type weight), score is calculated as 1 / idx1 / idx2 for each item & candidate pair, then sum the score as candidate score\n 2. co-visit features for rerank model - counts & probs pairwise item co-occurrences, index distance & time distance(consider order or not), etc.\n- **co-visit2(buy2buy)** - only carts or orders\n- **co-visit3(next item)** - only consider last item when generate candidate\n- **sknn1** - simple sknn with 100 most similair & recent session \n- **sknn2** - a variant of sknn named as Sequence and Time Aware Neighborhood, reference https://arxiv.org/abs/1910.12781 for both sknn implementation\n\n###### Re-Rank Model\n\n- **aid features** - simple stat by time window(7, 14, 21)\n- **session features** - length, aid count, count by type, time to the end of prediction period\n- **revisit aid X session**, clicks/carts/orders count, absolute/relative position\n- **co-visit features** - as mentioned before, generate a lot feature to describe pairwise item co-occurrences\n- **sknn features** - only candidate score/rank\n- **similarity features** - similarity between candidate item and session by item embedding of word2vec & implicit' s BPR\n\nFinally, 3 models(clicks/carts/orders) is trained by LightGBM(lambdarank)", "votes": null }, { "id": "2124886", "postDate": "02/01/2023 09:23:15", "content": "Congrats for solo gold!
", "rawMarkdown": "Congrats for solo gold!", "votes": null }, { "id": "2125043", "postDate": "02/01/2023 11:37:03", "content": "Congratulation!👍 I have a question about similarity features. Do you present the session embedding by using the average of the historical aid sequence? or else.
", "rawMarkdown": "Congratulation!👍 I have a question about similarity features. Do you present the session embedding by using the average of the historical aid sequence? or else.", "votes": null }, { "id": "2125051", "postDate": "02/01/2023 11:43:23", "content": "Thanks. I calculate the cosine distance between the candidate and last item, and max, min, mean to each item in the session.
", "rawMarkdown": "Thanks. I calculate the cosine distance between the candidate and last item, and max, min, mean to each item in the session.", "votes": null }, { "id": "2125061", "postDate": "02/01/2023 11:56:09", "content": "Thanks, lucky solo, let's find a chance to team up.
", "rawMarkdown": "Thanks, lucky solo, let's find a chance to team up.", "votes": null }, { "id": "2125113", "postDate": "02/01/2023 12:38:46", "content": "You didn't take only 20/50/last day features, used all aids in session?
", "rawMarkdown": "You didn't take only 20/50/last day features, used all aids in session?", "votes": null }, { "id": "2125124", "postDate": "02/01/2023 12:48:46", "content": "Hi @gongbi congrats for solo gold!
\none question, which library do you use for sknn?
", "rawMarkdown": "Hi @gongbi congrats for solo gold!\n\none question, which library do you use for sknn?", "votes": null }, { "id": "2125281", "postDate": "02/01/2023 15:12:32", "content": "just implement it by myself, the logic is simple.
", "rawMarkdown": "just implement it by myself, the logic is simple.", "votes": null }, { "id": "2125308", "postDate": "02/01/2023 15:28:14", "content": "Congras and thanks for sharing. May I ask what training data and label are you using ?? RADEK's CV data here -> https://www.kaggle.com/competitions/otto-recommender-system/discussion/364991 or you use the whole train and test data.
", "rawMarkdown": "Congras and thanks for sharing. May I ask what training data and label are you using ?? RADEK's CV data here -> https://www.kaggle.com/competitions/otto-recommender-system/discussion/364991 or you use the whole train and test data.", "votes": null }, { "id": "2125323", "postDate": "02/01/2023 15:34:59", "content": "I built my own code to prepare the training data and label, logic should be same.
", "rawMarkdown": "I built my own code to prepare the training data and label, logic should be same.", "votes": null }, { "id": "2125396", "postDate": "02/01/2023 16:22:20", "content": "Yes, i only take last 40 aid, not selected carefully. And I have another group of feature to take carts or orders only.
", "rawMarkdown": "Yes, i only take last 40 aid, not selected carefully. And I have another group of feature to take carts or orders only.", "votes": null }, { "id": "2126127", "postDate": "02/02/2023 05:33:13", "content": "Congrats on your solo gold! Amazing work!
\nThe Devastator.
", "rawMarkdown": "Congrats on your solo gold! Amazing work!\n\nThe Devastator.", "votes": null }, { "id": "2126802", "postDate": "02/02/2023 13:07:41", "content": "What is sknn? What is his full name?
", "rawMarkdown": "What is sknn? What is his full name?", "votes": null }, { "id": "2126814", "postDate": "02/02/2023 13:20:24", "content": "Noted, thanks for your reply
", "rawMarkdown": "Noted, thanks for your reply", "votes": null } ], "comments": [ { "id": 2124886, "author_name": "chenxin1991", "author_url": "", "post_date": "02/01/2023 09:23:15", "content": "Congrats for solo gold!
", "votes": null, "replies": [ { "id": 2125061, "author_name": "gongbi", "author_url": "", "post_date": "02/01/2023 11:56:09", "content": "Thanks, lucky solo, let's find a chance to team up.
", "votes": null, "replies": [] } ] }, { "id": 2125043, "author_name": "deeeeeeeplearning", "author_url": "", "post_date": "02/01/2023 11:37:03", "content": "Congratulation!👍 I have a question about similarity features. Do you present the session embedding by using the average of the historical aid sequence? or else.
", "votes": null, "replies": [ { "id": 2125051, "author_name": "gongbi", "author_url": "", "post_date": "02/01/2023 11:43:23", "content": "Thanks. I calculate the cosine distance between the candidate and last item, and max, min, mean to each item in the session.
", "votes": null, "replies": [ { "id": 2125113, "author_name": "artemfedorov", "author_url": "", "post_date": "02/01/2023 12:38:46", "content": "You didn't take only 20/50/last day features, used all aids in session?
", "votes": null, "replies": [ { "id": 2125396, "author_name": "gongbi", "author_url": "", "post_date": "02/01/2023 16:22:20", "content": "Yes, i only take last 40 aid, not selected carefully. And I have another group of feature to take carts or orders only.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2125124, "author_name": "trasibulo", "author_url": "", "post_date": "02/01/2023 12:48:46", "content": "Hi @gongbi congrats for solo gold!
\none question, which library do you use for sknn?
", "votes": null, "replies": [ { "id": 2125281, "author_name": "gongbi", "author_url": "", "post_date": "02/01/2023 15:12:32", "content": "just implement it by myself, the logic is simple.
", "votes": null, "replies": [] } ] }, { "id": 2125308, "author_name": "weiqiu", "author_url": "", "post_date": "02/01/2023 15:28:14", "content": "Congras and thanks for sharing. May I ask what training data and label are you using ?? RADEK's CV data here -> https://www.kaggle.com/competitions/otto-recommender-system/discussion/364991 or you use the whole train and test data.
", "votes": null, "replies": [ { "id": 2125323, "author_name": "gongbi", "author_url": "", "post_date": "02/01/2023 15:34:59", "content": "I built my own code to prepare the training data and label, logic should be same.
", "votes": null, "replies": [ { "id": 2126814, "author_name": "weiqiu", "author_url": "", "post_date": "02/02/2023 13:20:24", "content": "Noted, thanks for your reply
", "votes": null, "replies": [] } ] } ] }, { "id": 2126127, "author_name": "thedevastator", "author_url": "", "post_date": "02/02/2023 05:33:13", "content": "Congrats on your solo gold! Amazing work!
\nThe Devastator.
", "votes": null, "replies": [] }, { "id": 2126802, "author_name": "yasso1", "author_url": "", "post_date": "02/02/2023 13:07:41", "content": "What is sknn? What is his full name?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2124826": "Thanks OTTO & Kaggle to host such interesting competition, my solution is straight forward, just candidate generation + reranker, a short summary as below.\n\n###### Candidate Generation\n\n- **re-visit** - all visited items\n- **co-visit1** - based on the public notebook with 2 improvement. Thanks @radek1 and @cdeotte \n 1. candidate generation by weight decay, e.g. a session has N items(idx1 by the reverse ts) & each item has M pairs of co-visit item(idx2 by time weight or type weight), score is calculated as 1 / idx1 / idx2 for each item & candidate pair, then sum the score as candidate score\n 2. co-visit features for rerank model - counts & probs pairwise item co-occurrences, index distance & time distance(consider order or not), etc.\n- **co-visit2(buy2buy)** - only carts or orders\n- **co-visit3(next item)** - only consider last item when generate candidate\n- **sknn1** - simple sknn with 100 most similair & recent session \n- **sknn2** - a variant of sknn named as Sequence and Time Aware Neighborhood, reference https://arxiv.org/abs/1910.12781 for both sknn implementation\n\n###### Re-Rank Model\n\n- **aid features** - simple stat by time window(7, 14, 21)\n- **session features** - length, aid count, count by type, time to the end of prediction period\n- **revisit aid X session**, clicks/carts/orders count, absolute/relative position\n- **co-visit features** - as mentioned before, generate a lot feature to describe pairwise item co-occurrences\n- **sknn features** - only candidate score/rank\n- **similarity features** - similarity between candidate item and session by item embedding of word2vec & implicit' s BPR\n\nFinally, 3 models(clicks/carts/orders) is trained by LightGBM(lambdarank)", "2124886": "Congrats for solo gold!", "2125043": "Congratulation!👍 I have a question about similarity features. Do you present the session embedding by using the average of the historical aid sequence? or else.", "2125051": "Thanks. I calculate the cosine distance between the candidate and last item, and max, min, mean to each item in the session.", "2125061": "Thanks, lucky solo, let's find a chance to team up.", "2125113": "You didn't take only 20/50/last day features, used all aids in session?", "2125124": "Hi @gongbi congrats for solo gold!\n\none question, which library do you use for sknn?", "2125281": "just implement it by myself, the logic is simple.", "2125308": "Congras and thanks for sharing. May I ask what training data and label are you using ?? RADEK's CV data here -> https://www.kaggle.com/competitions/otto-recommender-system/discussion/364991 or you use the whole train and test data.", "2125323": "I built my own code to prepare the training data and label, logic should be same.", "2125396": "Yes, i only take last 40 aid, not selected carefully. And I have another group of feature to take carts or orders only.", "2126127": "Congrats on your solo gold! Amazing work!\n\nThe Devastator.", "2126802": "What is sknn? What is his full name?", "2126814": "Noted, thanks for your reply" }, "source": "meta" }