\nThis is my overall model framework.
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F921191%2F7164310586891a03ca0ff7a8a86b6422%2Fpipline_20230206230744.png?generation=1675696578724236&alt=media\")
Thanks to kaggle and OTTO for the great game. This is my first solo gold medal and I'm very excited about it.
\nThis is my overall model framework.
\n
I've included three recalls
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F921191%2Ffffea55d055186ec905fe62678dd819e%2Fi2i_sim_20230207001335.png?generation=1675700054785631&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F921191%2F73e54cb63b4c969f0091a271e140b481%2Ftrain_valid_20230206232247.png?generation=1675697003585741&alt=media\")
\nIn the verification phase, I used Radek’s CV strategy.
\nFor online prediction, train_v1 + train_v2 + valid was used as training data.
I used lightgbm binary classifier, learning rate: 0.02, 5500 rounds
\nbest single model local cv parts:
\n1.orders recall@20 is 0.6715
\n1.carts recall@20 is 0.4433
\n1.clicks recall@20 is 0.5561
\nMy local cv is 0.6715 * 0.6 + 0.4433 * 0.3 + 0.5561 * 0.1 = 0.5915
\nLB is 0.60335
I didn't run the second model，i am using the previously submitted version of the model for a blend of probability. this is give me the final score 0.60341
", "messages": [ { "id": 2132260, "postDate": "2023-02-06T17:33:59.683Z", "content": "Thanks to kaggle and OTTO for the great game. This is my first solo gold medal and I'm very excited about it.
\nThis is my overall model framework.
\n
I've included three recalls
\n
\nIn the verification phase, I used Radek’s CV strategy.
\nFor online prediction, train_v1 + train_v2 + valid was used as training data.
I used lightgbm binary classifier, learning rate: 0.02, 5500 rounds
\nbest single model local cv parts:
\n1.orders recall@20 is 0.6715
\n1.carts recall@20 is 0.4433
\n1.clicks recall@20 is 0.5561
\nMy local cv is 0.6715 * 0.6 + 0.4433 * 0.3 + 0.5561 * 0.1 = 0.5915
\nLB is 0.60335
I didn't run the second model，i am using the previously submitted version of the model for a blend of probability. this is give me the final score 0.60341
", "rawMarkdown": "Thanks to kaggle and OTTO for the great game. This is my first solo gold medal and I'm very excited about it.\nThis is my overall model framework.\n\n\n# Retrieval\nI've included three recalls\n- top 150 Co-visitation Matrix by CHRIS DEOTTE [https://www.kaggle.com/code/cdeotte/candidate-rerank-model-lb-0-575](url)\n- top 100 click 2 click bidirection i2i similarity with pos, time, session, aid weight \n- top 100 click 2 cart bidirection i2i similarity with pos, time, session, aid weight \n\n\n# Features\n- session feats: \n1. the counts and frequency of user clicks/orders/carts\n1. user last clicks/orders/carts aid and hour\n1. user last behavior type\n- aid feats:\n1. aid clicks/orders/carts counts\n1. aid clicks/orders/carts ratio\n1. aid clicks/orders/carts time\n1. aid behavior mean type\n- session aid feats:\n1. user clicks/orders/carts aid counts\n1. user clicks/orders/carts aid time\n1. user behavior aid mean type and last behavior type\n1. abs(hots/time of user click/cart/order the aid - aid click/cart/order hots/time)\n- sim feats:\n1. Co-visitation Matrix rank\n1. clicks/carts/orders to clicks/carts/orders 2 clicks/carts/orders i2i/i2i2i sim with pos, time weight.\n1. clicks/carts/orders to clicks/carts/orders 2 clicks/carts/orders i2i/i2i2i mean/max/min/std/last sim \n1. clicks/carts/orders to clicks/carts/orders aid pair sim with pos, time weight.\n1. clicks/carts/orders to clicks/carts/orders aid pair mean/max/min/std/last sim \n1. w2v embedding sim \n\n# Train And Validation \n\nIn the verification phase, I used Radek’s CV strategy.\nFor online prediction, train_v1 + train_v2 + valid was used as training data.\n\n# Model\nI used lightgbm binary classifier, learning rate: 0.02, 5500 rounds\n\n# Local CV And LB Score\nbest single model local cv parts:\n1.orders recall@20 is 0.6715\n1.carts recall@20 is 0.4433\n1.clicks recall@20 is 0.5561\nMy local cv is 0.6715 * 0.6 + 0.4433 * 0.3 + 0.5561 * 0.1 = 0.5915\nLB is 0.60335\n\n# Enemble\nI didn't run the second model，i am using the previously submitted version of the model for a blend of probability. this is give me the final score 0.60341", "votes": 44 }, { "id": 2159987, "postDate": "2023-02-26T08:50:08.783Z", "content": "excuse me, may i ask what is the score that only i2i recall, not rank and not 0.575 can you get
", "rawMarkdown": "excuse me, may i ask what is the score that only i2i recall, not rank and not 0.575 can you get" }, { "id": 2156020, "postDate": "2023-02-23T03:51:46.980Z", "content": "Thank you so much for this!
\nIf it doesn't bother you, may I ask what is position order of an item, and is_sort value, which was mentioned in the i2i similarity?
Thank you so much for this. Really helpful
", "rawMarkdown": "Thank you so much for this. Really helpful " }, { "id": 2134462, "postDate": "2023-02-08T02:08:15.937Z", "content": "@ilovearsenal The leaderboard is finalized and you are 7th place now. Do you modify the title of thread ? We are 8th place now.
", "rawMarkdown": "@ilovearsenal The leaderboard is finalized and you are 7th place now. Do you modify the title of thread ? We are 8th place now.", "replies": [ { "id": 2134643, "postDate": "2023-02-08T06:52:16.787Z", "content": "ok,i have modified the title😄.
", "rawMarkdown": "ok,i have modified the title😄.", "votes": 1 } ] }, { "id": 2132658, "postDate": "2023-02-07T01:00:12.723Z", "content": "Congrats and thanks for sharing!!
\nI also tried i2i similarity but ultimately gave up due to memory issues. I wonder how did you manage memory when calculating i2i similarity or you just had better hardware specs?
Thank you for your question, I used a machine with 128G RAM and 16 core cpu.
\nI used dict to store similarity data.
\nThere are two limitations to calculating i2i similarity.
\n1.Collinear relationships are calculated only for N days.
\n2.Collinear relationships are calculated only for position distance <=M.
excuse me, may i ask what is the score that only i2i recall, not rank and not 0.575 can you get
", "votes": 0, "replies": [] }, { "id": 2156020, "author_name": "Ngọc Uyên Nguyễn", "author_url": "", "post_date": "2023-02-23T03:51:46.980000", "content": "Thank you so much for this!
\nIf it doesn't bother you, may I ask what is position order of an item, and is_sort value, which was mentioned in the i2i similarity?
Thank you so much for this. Really helpful
", "votes": 0, "replies": [] }, { "id": 2134462, "author_name": "toshi_k", "author_url": "", "post_date": "2023-02-08T02:08:15.937000", "content": "@ilovearsenal The leaderboard is finalized and you are 7th place now. Do you modify the title of thread ? We are 8th place now.
", "votes": 0, "replies": [ { "id": 2134643, "author_name": "THLUO", "author_url": "", "post_date": "2023-02-08T06:52:16.787000", "content": "ok,i have modified the title😄.
", "votes": 1, "replies": [] } ] }, { "id": 2132658, "author_name": "Bibek", "author_url": "", "post_date": "2023-02-07T01:00:12.723000", "content": "Congrats and thanks for sharing!!
\nI also tried i2i similarity but ultimately gave up due to memory issues. I wonder how did you manage memory when calculating i2i similarity or you just had better hardware specs?
Thank you for your question, I used a machine with 128G RAM and 16 core cpu.
\nI used dict to store similarity data.
\nThere are two limitations to calculating i2i similarity.
\n1.Collinear relationships are calculated only for N days.
\n2.Collinear relationships are calculated only for position distance <=M.