Congrats to all you put best efforts for your medal standings .All those who couldn't get desired medal this time, just learn and move on for better next time .
\nOurs final standing was 24th ,few last things at finish line helped us push a bit .
\nMy team members were @ekffar @cswwp347724 they were fabulous worked hard till end .
\nOur score was powered by @cswwp347724 sampling strategy ,saint plus , critical bug solving related to lagtime towards end by me, some feature addition related to lecture, ensembling of lgb with saint plus folds 1, 3 by @ekffar well done guys both of you .
**For saint plus most important part Calculation Lagtime. We tried prior lag time as in paper and simple lag time ts2-ts1, second one turned out to be the best.
\n**
\nMost tricky part of it was setting it right during inference. We should be considering lagtime for test q coming in next iterations with respect to last task container in previous iterations. Many people might have missed this so their Saint Plus score could have got stuck at 79.2-3. I identified this mistake and immediately our score boosted by 0.005.
Detailed solution is here
\nThanks @cswwp347724 for putting one in place.
\nhttps://www.kaggle.com/c/riiid-test-answer-prediction/discussion/209659
", "messages": [ { "id": "1143766", "postDate": "01/08/2021 03:53:34", "content": "Congrats to all you put best efforts for your medal standings .All those who couldn't get desired medal this time, just learn and move on for better next time .
\nOurs final standing was 24th ,few last things at finish line helped us push a bit .
\nMy team members were @ekffar @cswwp347724 they were fabulous worked hard till end .
\nOur score was powered by @cswwp347724 sampling strategy ,saint plus , critical bug solving related to lagtime towards end by me, some feature addition related to lecture, ensembling of lgb with saint plus folds 1, 3 by @ekffar well done guys both of you .
**For saint plus most important part Calculation Lagtime. We tried prior lag time as in paper and simple lag time ts2-ts1, second one turned out to be the best.
\n**
\nMost tricky part of it was setting it right during inference. We should be considering lagtime for test q coming in next iterations with respect to last task container in previous iterations. Many people might have missed this so their Saint Plus score could have got stuck at 79.2-3. I identified this mistake and immediately our score boosted by 0.005.
Detailed solution is here
\nThanks @cswwp347724 for putting one in place.
\nhttps://www.kaggle.com/c/riiid-test-answer-prediction/discussion/209659
", "rawMarkdown": "Congrats to all you put best efforts for your medal standings .All those who couldn't get desired medal this time, just learn and move on for better next time .\n\nOurs final standing was 24th ,few last things at finish line helped us push a bit .\nMy team members were @ekffar @cswwp347724 they were fabulous worked hard till end .\nOur score was powered by @cswwp347724 sampling strategy ,saint plus , critical bug solving related to lagtime towards end by me, some feature addition related to lecture, ensembling of lgb with saint plus folds 1, 3 by @ekffar well done guys both of you .\n\n**For saint plus most important part Calculation Lagtime. We tried prior lag time as in paper and simple lag time ts2-ts1, second one turned out to be the best. \n**\nMost tricky part of it was setting it right during inference. We should be considering lagtime for test q coming in next iterations with respect to last task container in previous iterations. Many people might have missed this so their Saint Plus score could have got stuck at 79.2-3. I identified this mistake and immediately our score boosted by 0.005. \n\nDetailed solution is here \n\nThanks @cswwp347724 for putting one in place.\n\n\nhttps://www.kaggle.com/c/riiid-test-answer-prediction/discussion/209659", "votes": null }, { "id": "1143767", "postDate": "01/08/2021 03:55:10", "content": "Congratulations Team. What lecture features did you use?
", "rawMarkdown": "Congratulations Team. What lecture features did you use?", "votes": null }, { "id": "1143773", "postDate": "01/08/2021 04:00:55", "content": "@cdeotte congrats for your team gold medal.
\n This was very tricky at inference ,had to think of multiple scenarios to fit this feature right helped by ekffar in building those .
\nIts Prior interaction as lecture or not .
\nI planned to add some more to it like lecture elapsed time and lectures part ,but couldn't get time to work and test on it .At first attempt we added prior interaction lecture or not that gave a push of 0.002 to cv ,then I thought would add l elt and its part info also thought process was may be model can correlate students ability to solve parts if he went through associated lecture in prior interaction , but this pushed cv down so we kept only lecture or not feature .I guess some more experiments with lecture features could have got more but lack of time .
@jaideepvalani great job. You folks narrowly missed the gold..
", "rawMarkdown": "jaideepvalani great job. You folks narrowly missed the gold..", "votes": null }, { "id": "1143796", "postDate": "01/08/2021 04:24:07", "content": "Yeah but Gold is always a gold
", "rawMarkdown": "Yeah but Gold is always a gold", "votes": null }, { "id": "1143830", "postDate": "01/08/2021 04:52:52", "content": "Well done and congratulations to you and your team for your efforts.
", "rawMarkdown": "Well done and congratulations to you and your team for your efforts.", "votes": null }, { "id": "1143862", "postDate": "01/08/2021 05:31:14", "content": "Great Accomplishment!
", "rawMarkdown": "Great Accomplishment!", "votes": null }, { "id": "1144162", "postDate": "01/08/2021 09:17:58", "content": "Nice, thanks to my teammates @jaideepvalani @ekffar
", "rawMarkdown": "Nice, thanks to my teammates @jaideepvalani @ekffar", "votes": null } ], "comments": [ { "id": 1143767, "author_name": "cdeotte", "author_url": "", "post_date": "01/08/2021 03:55:10", "content": "Congratulations Team. What lecture features did you use?
", "votes": null, "replies": [ { "id": 1143773, "author_name": "jaideepvalani", "author_url": "", "post_date": "01/08/2021 04:00:55", "content": "@cdeotte congrats for your team gold medal.
\n This was very tricky at inference ,had to think of multiple scenarios to fit this feature right helped by ekffar in building those .
\nIts Prior interaction as lecture or not .
\nI planned to add some more to it like lecture elapsed time and lectures part ,but couldn't get time to work and test on it .At first attempt we added prior interaction lecture or not that gave a push of 0.002 to cv ,then I thought would add l elt and its part info also thought process was may be model can correlate students ability to solve parts if he went through associated lecture in prior interaction , but this pushed cv down so we kept only lecture or not feature .I guess some more experiments with lecture features could have got more but lack of time .
@jaideepvalani great job. You folks narrowly missed the gold..
", "votes": null, "replies": [ { "id": 1143796, "author_name": "jaideepvalani", "author_url": "", "post_date": "01/08/2021 04:24:07", "content": "Yeah but Gold is always a gold
", "votes": null, "replies": [] } ] }, { "id": 1143830, "author_name": "authman", "author_url": "", "post_date": "01/08/2021 04:52:52", "content": "Well done and congratulations to you and your team for your efforts.
", "votes": null, "replies": [] }, { "id": 1143862, "author_name": "adityaecdrid", "author_url": "", "post_date": "01/08/2021 05:31:14", "content": "Great Accomplishment!
", "votes": null, "replies": [] }, { "id": 1144162, "author_name": "cswwp347724", "author_url": "", "post_date": "01/08/2021 09:17:58", "content": "Nice, thanks to my teammates @jaideepvalani @ekffar
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1143766": "Congrats to all you put best efforts for your medal standings .All those who couldn't get desired medal this time, just learn and move on for better next time .\n\nOurs final standing was 24th ,few last things at finish line helped us push a bit .\nMy team members were @ekffar @cswwp347724 they were fabulous worked hard till end .\nOur score was powered by @cswwp347724 sampling strategy ,saint plus , critical bug solving related to lagtime towards end by me, some feature addition related to lecture, ensembling of lgb with saint plus folds 1, 3 by @ekffar well done guys both of you .\n\n**For saint plus most important part Calculation Lagtime. We tried prior lag time as in paper and simple lag time ts2-ts1, second one turned out to be the best. \n**\nMost tricky part of it was setting it right during inference. We should be considering lagtime for test q coming in next iterations with respect to last task container in previous iterations. Many people might have missed this so their Saint Plus score could have got stuck at 79.2-3. I identified this mistake and immediately our score boosted by 0.005. \n\nDetailed solution is here \n\nThanks @cswwp347724 for putting one in place.\n\n\nhttps://www.kaggle.com/c/riiid-test-answer-prediction/discussion/209659", "1143767": "Congratulations Team. What lecture features did you use?", "1143773": "cdeotte congrats for your team gold medal.\n This was very tricky at inference ,had to think of multiple scenarios to fit this feature right helped by ekffar in building those .\nIts Prior interaction as lecture or not .\nI planned to add some more to it like lecture elapsed time and lectures part ,but couldn't get time to work and test on it .At first attempt we added prior interaction lecture or not that gave a push of 0.002 to cv ,then I thought would add l elt and its part info also thought process was may be model can correlate students ability to solve parts if he went through associated lecture in prior interaction , but this pushed cv down so we kept only lecture or not feature .I guess some more experiments with lecture features could have got more but lack of time .", "1143782": "jaideepvalani great job. You folks narrowly missed the gold..", "1143796": "Yeah but Gold is always a gold", "1143830": "Well done and congratulations to you and your team for your efforts.", "1143862": "Great Accomplishment!", "1144162": "Nice, thanks to my teammates @jaideepvalani @ekffar" }, "source": "meta" }