I'm new here, and really worried about the incoming shake up. \nNow I have a model with best CV and best LB,then I simply blend it with the second good model and got a same score.\nI'm looking for a conservative strategy for final submission,so does anyone has some experience for final submission? \nby the way,in this competition,I wouldn't be surprised that when I wake up and find myself dropping out of the medal areas after visiting many shake up happened in previous competitions.XD
", "messages": [ { "id": 541457, "postDate": "2019-06-02T13:40:05.647Z", "content": "I'm new here, and really worried about the incoming shake up. \nNow I have a model with best CV and best LB,then I simply blend it with the second good model and got a same score.\nI'm looking for a conservative strategy for final submission,so does anyone has some experience for final submission? \nby the way,in this competition,I wouldn't be surprised that when I wake up and find myself dropping out of the medal areas after visiting many shake up happened in previous competitions.XD
", "rawMarkdown": "I'm new here, and really worried about the incoming shake up. \nNow I have a model with best CV and best LB,then I simply blend it with the second good model and got a same score.\nI'm looking for a conservative strategy for final submission,so does anyone has some experience for final submission? \nby the way,in this competition,I wouldn't be surprised that when I wake up and find myself dropping out of the medal areas after visiting many shake up happened in previous competitions.XD", "votes": 3 }, { "id": 541572, "postDate": "2019-06-02T16:58:26.690Z", "content": "for me who am in the first experience of competitions, these are my parameters to select my submission:
\n\n1 - CV / LB correlation\n2 - score in the public LB (if obtained with a succession of submissions with decreasing CV)\n3 - possibility to learn from my mistakes (no mixtures with public solutions or similar)\n4 - (last but not least) listen to the advice of the masters :-)
\n\nI have for my best CV/LB (1375) submission:
\n\nmean: 5.181616993559135\nmedian: 4.789336818679703\nstd: 2.605603236706993
", "rawMarkdown": "for me who am in the first experience of competitions, these are my parameters to select my submission:\n\n1 - CV / LB correlation\n2 - score in the public LB (if obtained with a succession of submissions with decreasing CV)\n3 - possibility to learn from my mistakes (no mixtures with public solutions or similar)\n4 - (last but not least) listen to the advice of the masters :-)\n\nI have for my best CV/LB (1375) submission:\n\nmean: 5.181616993559135\nmedian: 4.789336818679703\nstd: 2.605603236706993", "votes": 1 }, { "id": 541834, "postDate": "2019-06-03T04:52:08.417Z", "content": "How to avoid shake up.\nI have the similar topic, perhaps you can find the answer in the comments of my topic. (I've got my answer.
", "rawMarkdown": "[How to avoid shake up.](https://www.kaggle.com/c/LANL-Earthquake-Prediction/discussion/93679)\nI have the similar topic, perhaps you can find the answer in the comments of my topic. (I've got my answer.", "votes": 2 }, { "id": 541483, "postDate": "2019-06-02T14:17:40.813Z", "content": "We will use our CV score to select submissions. We discussed at length our validation strategy weeks ago, and now we're about to execute it. Unless people were on the moon during last few weeks, I'm sure they've read me saying that public LB is of not help in selecting final submission.
", "rawMarkdown": "We will use our CV score to select submissions. We discussed at length our validation strategy weeks ago, and now we're about to execute it. Unless people were on the moon during last few weeks, I'm sure they've read me saying that public LB is of not help in selecting final submission.", "votes": 2, "replies": [ { "id": 541494, "postDate": "2019-06-02T14:45:04.913Z", "content": "but you have 2 submissions.So you won't choose your highest LB?it's ok if you dont answer me ;)
", "rawMarkdown": "but you have 2 submissions.So you won't choose your highest LB?it's ok if you dont answer me ;)" }, { "id": 541504, "postDate": "2019-06-02T15:04:48.997Z", "content": "We will definitely not chose our best public LB as we know it overfits.
", "rawMarkdown": "We will definitely not chose our best public LB as we know it overfits.\n" }, { "id": 541725, "postDate": "2019-06-03T00:41:13.217Z", "content": "may I ask more , will you choose two submissions with similar CV or different CV? :p
", "rawMarkdown": "may I ask more , will you choose two submissions with similar CV or different CV? :p" }, { "id": 541730, "postDate": "2019-06-03T00:59:32.873Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 541465, "postDate": "2019-06-02T13:45:59.377Z", "content": "We have equivalent public LB scores, and you are above me. May I ask what is the mean and median of your test submission? If you have higher values, you might not drop below me.
", "rawMarkdown": "We have equivalent public LB scores, and you are above me. May I ask what is the mean and median of your test submission? If you have higher values, you might not drop below me.", "votes": 2, "replies": [ { "id": 541467, "postDate": "2019-06-02T13:51:10.613Z", "content": "what is the benchmark mean and median if you may like to disclose. And i hope your gold sticks as you have been working since the beginning of the competition. wish you luck.
", "rawMarkdown": "what is the benchmark mean and median if you may like to disclose. And i hope your gold sticks as you have been working since the beginning of the competition. wish you luck.", "votes": 1 }, { "id": 541472, "postDate": "2019-06-02T14:01:06.320Z", "content": "Yes, I will (partly) disclose it now, and hope some generous people (especially at the top 100) can share theirs too, so we all know where we're at at this moment. I have test mean 5.5 and median 5.03. All I can say now is just having higher (both values) might be better for private LB, given 2 submissions with identical public LB.
", "rawMarkdown": "Yes, I will (partly) disclose it now, and hope some generous people (especially at the top 100) can share theirs too, so we all know where we're at at this moment. I have test mean 5.5 and median 5.03. All I can say now is just having higher (both values) might be better for private LB, given 2 submissions with identical public LB. ", "votes": 2 }, { "id": 541476, "postDate": "2019-06-02T14:05:33.890Z", "content": "thanks , Now i am braced to see my self in no medal zone
\n\n\n>>> df.mean()\ntime_to_failure 5.017047\n>>> df.median()\ntime_to_failure 4.621451\n\nUPDATE:
\n>>> df.mean()\ntime_to_failure 5.592439\n>>> df.median()\ntime_to_failure 5.124839\n\nbut LB score in last one is 1.402
I may be wrong. But even suppose I'm right, how do you say that while you don't know where the others are at? That's why, I really hope for people to share these stats here, so we would be more relieved of being dropped, or not too surprised of being elevated.
", "rawMarkdown": "I may be wrong. But even suppose I'm right, how do you say that while you don't know where the others are at? That's why, I really hope for people to share these stats here, so we would be more relieved of being dropped, or not too surprised of being elevated. ", "votes": 1 }, { "id": 541479, "postDate": "2019-06-02T14:13:55.580Z", "content": "May be you can create a discussion topic for people to share their score of mean and median. i don't see any harm in sharing that. this way we will have more data and you will get to validate your hypothesis after the shake up
", "rawMarkdown": "May be you can create a discussion topic for people to share their score of mean and median. i don't see any harm in sharing that. this way we will have more data and you will get to validate your hypothesis after the shake up ", "votes": 1 }, { "id": 541491, "postDate": "2019-06-02T14:42:02.297Z", "content": "mean:time_to_failure 5.409001\nmedian:time_to_failure 5.045274\nhowever,when i blend them ,mean decreased while median increased. with identical lb
", "rawMarkdown": "mean:time_to_failure 5.409001\nmedian:time_to_failure 5.045274\nhowever,when i blend them ,mean decreased while median increased. with identical lb", "votes": 2 }, { "id": 541524, "postDate": "2019-06-02T15:52:40.693Z", "content": "Thanks for this discussion. In my case - the higher the LB is, the lower the median and the mean are. At the moment I have:\n4.785360842 median\n5.425311235 mean
", "rawMarkdown": "Thanks for this discussion. In my case - the higher the LB is, the lower the median and the mean are. At the moment I have:\n4.785360842 median\n5.425311235\tmean", "votes": 2 }, { "id": 541547, "postDate": "2019-06-02T16:25:48.547Z", "content": "@khahuras would you like to tell little bit more about your hypothesis that how are you arriving at this mean and median thing.
", "rawMarkdown": "@khahuras would you like to tell little bit more about your hypothesis that how are you arriving at this mean and median thing." }, { "id": 541734, "postDate": "2019-06-03T01:04:25.530Z", "content": "I may be wrong, so just let's wait for the shake to see...
", "rawMarkdown": "I may be wrong, so just let's wait for the shake to see..." } ] }, { "id": 542497, "postDate": "2019-06-04T00:28:54.457Z", "content": "Public LB sample size was too small. It really didn't mean anything.
", "rawMarkdown": "Public LB sample size was too small. It really didn't mean anything. " }, { "id": 541731, "postDate": "2019-06-03T01:01:06.977Z", "content": "The submission which use model that didn't rely on mean, or percentile mean, and will be best to predict EQ with TTL peak > 10. Thats my opinion. Thx.
", "rawMarkdown": "The submission which use model that didn't rely on mean, or percentile mean, and will be best to predict EQ with TTL peak > 10. Thats my opinion. Thx. " }, { "id": 541519, "postDate": "2019-06-02T15:49:20.400Z", "rawMarkdown": "", "votes": 1, "isDeleted": true, "replies": [ { "id": 541523, "postDate": "2019-06-02T15:50:33.217Z", "content": "Higher = better or worse?
", "rawMarkdown": "Higher = better or worse?" }, { "id": 541525, "postDate": "2019-06-02T15:53:47.283Z", "content": "the higher = better
", "rawMarkdown": "the higher = better " } ] } ], "comments": [ { "id": 541572, "author_name": "Bellò Filippo", "author_url": "", "post_date": "2019-06-02T16:58:26.690000", "content": "for me who am in the first experience of competitions, these are my parameters to select my submission:
\n\n1 - CV / LB correlation\n2 - score in the public LB (if obtained with a succession of submissions with decreasing CV)\n3 - possibility to learn from my mistakes (no mixtures with public solutions or similar)\n4 - (last but not least) listen to the advice of the masters :-)
\n\nI have for my best CV/LB (1375) submission:
\n\nmean: 5.181616993559135\nmedian: 4.789336818679703\nstd: 2.605603236706993
", "votes": 1, "replies": [] }, { "id": 541834, "author_name": "Zhongkai Shangguan", "author_url": "", "post_date": "2019-06-03T04:52:08.417000", "content": "How to avoid shake up.\nI have the similar topic, perhaps you can find the answer in the comments of my topic. (I've got my answer.
", "votes": 2, "replies": [] }, { "id": 541483, "author_name": "CPMP", "author_url": "", "post_date": "2019-06-02T14:17:40.813000", "content": "We will use our CV score to select submissions. We discussed at length our validation strategy weeks ago, and now we're about to execute it. Unless people were on the moon during last few weeks, I'm sure they've read me saying that public LB is of not help in selecting final submission.
", "votes": 2, "replies": [ { "id": 541494, "author_name": "ZeroWen", "author_url": "", "post_date": "2019-06-02T14:45:04.913000", "content": "but you have 2 submissions.So you won't choose your highest LB?it's ok if you dont answer me ;)
", "votes": 0, "replies": [] }, { "id": 541504, "author_name": "CPMP", "author_url": "", "post_date": "2019-06-02T15:04:48.997000", "content": "We will definitely not chose our best public LB as we know it overfits.
", "votes": 0, "replies": [] }, { "id": 541725, "author_name": "ZeroWen", "author_url": "", "post_date": "2019-06-03T00:41:13.217000", "content": "may I ask more , will you choose two submissions with similar CV or different CV? :p
", "votes": 0, "replies": [] }, { "id": 541730, "author_name": "", "author_url": "", "post_date": "2019-06-03T00:59:32.873000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 541465, "author_name": "Kha Vo", "author_url": "", "post_date": "2019-06-02T13:45:59.377000", "content": "We have equivalent public LB scores, and you are above me. May I ask what is the mean and median of your test submission? If you have higher values, you might not drop below me.
", "votes": 2, "replies": [ { "id": 541467, "author_name": "Manoj", "author_url": "", "post_date": "2019-06-02T13:51:10.613000", "content": "what is the benchmark mean and median if you may like to disclose. And i hope your gold sticks as you have been working since the beginning of the competition. wish you luck.
", "votes": 1, "replies": [] }, { "id": 541472, "author_name": "Kha Vo", "author_url": "", "post_date": "2019-06-02T14:01:06.320000", "content": "Yes, I will (partly) disclose it now, and hope some generous people (especially at the top 100) can share theirs too, so we all know where we're at at this moment. I have test mean 5.5 and median 5.03. All I can say now is just having higher (both values) might be better for private LB, given 2 submissions with identical public LB.
", "votes": 2, "replies": [] }, { "id": 541476, "author_name": "Manoj", "author_url": "", "post_date": "2019-06-02T14:05:33.890000", "content": "thanks , Now i am braced to see my self in no medal zone
\n\n\n>>> df.mean()\ntime_to_failure 5.017047\n>>> df.median()\ntime_to_failure 4.621451\n\nUPDATE:
\n>>> df.mean()\ntime_to_failure 5.592439\n>>> df.median()\ntime_to_failure 5.124839\n\nbut LB score in last one is 1.402
I may be wrong. But even suppose I'm right, how do you say that while you don't know where the others are at? That's why, I really hope for people to share these stats here, so we would be more relieved of being dropped, or not too surprised of being elevated.
", "votes": 1, "replies": [] }, { "id": 541479, "author_name": "Manoj", "author_url": "", "post_date": "2019-06-02T14:13:55.580000", "content": "May be you can create a discussion topic for people to share their score of mean and median. i don't see any harm in sharing that. this way we will have more data and you will get to validate your hypothesis after the shake up
", "votes": 1, "replies": [] }, { "id": 541491, "author_name": "ZeroWen", "author_url": "", "post_date": "2019-06-02T14:42:02.297000", "content": "mean:time_to_failure 5.409001\nmedian:time_to_failure 5.045274\nhowever,when i blend them ,mean decreased while median increased. with identical lb
", "votes": 2, "replies": [] }, { "id": 541524, "author_name": "AM", "author_url": "", "post_date": "2019-06-02T15:52:40.693000", "content": "Thanks for this discussion. In my case - the higher the LB is, the lower the median and the mean are. At the moment I have:\n4.785360842 median\n5.425311235 mean
", "votes": 2, "replies": [] }, { "id": 541547, "author_name": "Manoj", "author_url": "", "post_date": "2019-06-02T16:25:48.547000", "content": "@khahuras would you like to tell little bit more about your hypothesis that how are you arriving at this mean and median thing.
", "votes": 0, "replies": [] }, { "id": 541734, "author_name": "Kha Vo", "author_url": "", "post_date": "2019-06-03T01:04:25.530000", "content": "I may be wrong, so just let's wait for the shake to see...
", "votes": 0, "replies": [] } ] }, { "id": 542497, "author_name": "joejeo1", "author_url": "", "post_date": "2019-06-04T00:28:54.457000", "content": "Public LB sample size was too small. It really didn't mean anything.
", "votes": 0, "replies": [] }, { "id": 541731, "author_name": "OverfitModel", "author_url": "", "post_date": "2019-06-03T01:01:06.977000", "content": "The submission which use model that didn't rely on mean, or percentile mean, and will be best to predict EQ with TTL peak > 10. Thats my opinion. Thx.
", "votes": 0, "replies": [] }, { "id": 541519, "author_name": "", "author_url": "", "post_date": "2019-06-02T15:49:20.400000", "content": "", "votes": 1, "replies": [ { "id": 541523, "author_name": "Kha Vo", "author_url": "", "post_date": "2019-06-02T15:50:33.217000", "content": "Higher = better or worse?
", "votes": 0, "replies": [] }, { "id": 541525, "author_name": "AM", "author_url": "", "post_date": "2019-06-02T15:53:47.283000", "content": "the higher = better
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "541457": "I'm new here, and really worried about the incoming shake up. \nNow I have a model with best CV and best LB,then I simply blend it with the second good model and got a same score.\nI'm looking for a conservative strategy for final submission,so does anyone has some experience for final submission? \nby the way,in this competition,I wouldn't be surprised that when I wake up and find myself dropping out of the medal areas after visiting many shake up happened in previous competitions.XD", "541572": "for me who am in the first experience of competitions, these are my parameters to select my submission:\n\n1 - CV / LB correlation\n2 - score in the public LB (if obtained with a succession of submissions with decreasing CV)\n3 - possibility to learn from my mistakes (no mixtures with public solutions or similar)\n4 - (last but not least) listen to the advice of the masters :-)\n\nI have for my best CV/LB (1375) submission:\n\nmean: 5.181616993559135\nmedian: 4.789336818679703\nstd: 2.605603236706993", "541834": "[How to avoid shake up.](https://www.kaggle.com/c/LANL-Earthquake-Prediction/discussion/93679)\nI have the similar topic, perhaps you can find the answer in the comments of my topic. (I've got my answer.", "541483": "We will use our CV score to select submissions. We discussed at length our validation strategy weeks ago, and now we're about to execute it. Unless people were on the moon during last few weeks, I'm sure they've read me saying that public LB is of not help in selecting final submission.", "541465": "We have equivalent public LB scores, and you are above me. May I ask what is the mean and median of your test submission? If you have higher values, you might not drop below me.", "542497": "Public LB sample size was too small. It really didn't mean anything. ", "541731": "The submission which use model that didn't rely on mean, or percentile mean, and will be best to predict EQ with TTL peak > 10. Thats my opinion. Thx. ", "541519": "" } }