Just curious what is the best AUC and MCC folks were able to get to without using the magic features.
", "messages": [ { "id": 139100, "postDate": "2016-10-12T14:34:30.477Z", "content": "Using time correlations ( but no row order correlations or ID's) and some model based feature engineering is possible to push the performance to 0.0281 logLoss / 0.807 AUC/ 0.317 MCC (0.3166 LB)
\n\nAnyway, some "real" phenomena can cause row duplicates in the data: after a comunication crash, SCADA ( the subsystem responsible for reading and storing the measurements) usually repeat the last seen value, so the measurements associated to a given sequence of part numbers corresponds to the last correctly received, until the comunications is recovered.
\n\nSupose the comunication failure is produced by a power glitch ( or blackout) and this glitch also affect the parts quality currently manufactured. Then the first part number in a sequence of duplications will fail the quality check.
\n\nDoesn't sounds familiar to you? ;)
\n\nJust writing some science-fiction about "magic data properties"
\n\n"Any sufficiently advanced technology is indistinguishable from magic"
\n\nClarke's 3'rd law
", "rawMarkdown": "Using time correlations ( but no row order correlations or ID's) and some *model based* feature engineering is possible to push the performance to **0.0281 logLoss / 0.807 AUC/ 0.317 MCC (0.3166 LB)**\r\n\r\nAnyway, some \"real\" phenomena can cause row duplicates in the data: after a comunication crash, SCADA ( the subsystem responsible for reading and storing the measurements) usually repeat the last seen value, so the measurements associated to a given sequence of part numbers corresponds to the last correctly received, until the comunications is recovered.\r\n\r\nSupose the comunication failure is produced by a power glitch ( or blackout) and this glitch also affect the parts quality currently manufactured. Then the first part number in a sequence of duplications will fail the quality check.\r\n\r\nDoesn't sounds familiar to you? ;)\r\n\r\nJust writing some science-fiction about \"magic data properties\"\r\n\r\n*\"Any sufficiently advanced technology is indistinguishable from magic\"*\r\n\r\n*Clarke's 3'rd law*", "votes": 16 }, { "id": 139081, "postDate": "2016-10-12T13:17:06.047Z", "content": "Using only individual row information is hard to beat 0.31MCC and 0.78AUC and that model have a big value to Bosch. Using any kind of "data property feature" (there are many ways to explore it not using ID) it is possible to beat 0.5x, 0.93AUC. My current solution don't use IDs.
", "rawMarkdown": "Using only individual row information is hard to beat 0.31MCC and 0.78AUC and that model have a big value to Bosch. Using any kind of \"data property feature\" (there are many ways to explore it not using ID) it is possible to beat 0.5x, 0.93AUC. My current solution don't use IDs.", "votes": 9 }, { "id": 138922, "postDate": "2016-10-11T17:43:18.027Z", "content": "[quote=gingerman;138897]
\n\nJust edited my answer at the first sentence was a fragment - @anokas reported not to be using ID at all and it seems plausible to me.
\n\n[/quote]
\n\nYep, you are correct. We had 0.921 AUC before we added our first ID/row order feature.
", "rawMarkdown": "[quote=gingerman;138897]\r\n\r\nJust edited my answer at the first sentence was a fragment - @anokas reported not to be using ID at all and it seems plausible to me.\r\n\r\n[/quote]\r\n\r\nYep, you are correct. We had 0.921 AUC before we added our first ID/row order feature.", "votes": 1 }, { "id": 138897, "postDate": "2016-10-11T14:57:02.530Z", "content": "Just edited my answer at the first sentence was a fragment - @anokas reported not to be using ID at all and it seems plausible to me.
", "rawMarkdown": "Just edited my answer at the first sentence was a fragment - @anokas reported not to be using ID at all and it seems plausible to me.", "votes": 1 }, { "id": 138893, "postDate": "2016-10-11T14:31:37.117Z", "content": "I guess this cannot be answered as one can indirectly leverage the leakage without using IDs. Still, @anokas has reported ~0.92 AUC and ~0.49 MCC without magic feature.
", "rawMarkdown": "I guess this cannot be answered as one can indirectly leverage the leakage without using IDs. Still, @anokas has reported ~0.92 AUC and ~0.49 MCC without magic feature.", "votes": 1 }, { "id": 139462, "postDate": "2016-10-14T09:09:26.617Z", "content": "@Vicens,
\n\nyou just explained why the magic feature may have some meaningful, physical, reason to be there.
\n\nGreat insight.
", "rawMarkdown": "@Vicens,\r\n\r\nyou just explained why the magic feature may have some meaningful, physical, reason to be there.\r\n\r\nGreat insight.", "votes": 2 }, { "id": 140728, "postDate": "2016-10-22T16:05:16.860Z", "content": "@Jesse I am using 320 features only, including my engineered ones. Maybe I should use more ;)
\n\nIt would be great if an equivalent of Lasso is derived for GBM.
", "rawMarkdown": "@Jesse I am using 320 features only, including my engineered ones. Maybe I should use more ;)\r\n\r\nIt would be great if an equivalent of Lasso is derived for GBM. \r\n" }, { "id": 140696, "postDate": "2016-10-22T10:34:44.333Z", "content": "Jesse, there are various ways to capture that magic feature. Faron has shared one way, look for his kernel.
", "rawMarkdown": "Jesse, there are various ways to capture that magic feature. Faron has shared one way, look for his kernel." }, { "id": 139747, "postDate": "2016-10-16T09:05:35.343Z", "content": "Yes, there is higher probability but, clearly for now i don't see pattern for predicting the first, but there are patterns for second, third duplicates in a row. I am clearly missing something :)
", "rawMarkdown": "Yes, there is higher probability but, clearly for now i don't see pattern for predicting the first, but there are patterns for second, third duplicates in a row. I am clearly missing something :)" }, { "id": 139672, "postDate": "2016-10-15T18:49:42.617Z", "content": "@Salka
\nyes, specially those with consecutive Id's. And you can bet for the first in the row ;)
@Vicens Gaitan
\n\n\n\n\nUsing time correlations ( but no row order correlations or ID's)
\n
By this you mean when duplicates ocurr there is greater probability of some product to fail ?
", "rawMarkdown": "@Vicens Gaitan\r\n\r\n> Using time correlations ( but no row order correlations or ID's)\r\n\r\nBy this you mean when duplicates ocurr there is greater probability of some product to fail ?" }, { "id": 139594, "postDate": "2016-10-15T03:56:31.850Z", "content": "@Vicens, indeed thanks, very nice observations!
", "rawMarkdown": "@Vicens, indeed thanks, very nice observations!" }, { "id": 139076, "postDate": "2016-10-12T12:11:32.397Z", "content": "Hi raman,
\n\nIn this competition, there are some types of magical features except for ID.
\nI barely reach 0.30 without magical features.
Thanks! How about without using the ID feature at all? I am trying to see how generalizable the data models can be. That is, can we predict for a random object only using the non-id data.
", "rawMarkdown": "Thanks! How about without using the ID feature at all? I am trying to see how generalizable the data models can be. That is, can we predict for a random object only using the non-id data." }, { "id": 138890, "postDate": "2016-10-11T14:04:57.653Z", "content": "Just curious what is the best AUC and MCC folks were able to get to without using the magic features.
", "rawMarkdown": "Just curious what is the best AUC and MCC folks were able to get to without using the magic features." }, { "id": 140707, "postDate": "2016-10-22T12:19:57.327Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 140698, "postDate": "2016-10-22T10:37:59.387Z", "rawMarkdown": "", "isDeleted": true }, { "id": 140658, "postDate": "2016-10-22T01:42:20.773Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 139100, "author_name": "Vicens Gaitan", "author_url": "", "post_date": "2016-10-12T14:34:30.477000", "content": "Using time correlations ( but no row order correlations or ID's) and some model based feature engineering is possible to push the performance to 0.0281 logLoss / 0.807 AUC/ 0.317 MCC (0.3166 LB)
\n\nAnyway, some "real" phenomena can cause row duplicates in the data: after a comunication crash, SCADA ( the subsystem responsible for reading and storing the measurements) usually repeat the last seen value, so the measurements associated to a given sequence of part numbers corresponds to the last correctly received, until the comunications is recovered.
\n\nSupose the comunication failure is produced by a power glitch ( or blackout) and this glitch also affect the parts quality currently manufactured. Then the first part number in a sequence of duplications will fail the quality check.
\n\nDoesn't sounds familiar to you? ;)
\n\nJust writing some science-fiction about "magic data properties"
\n\n"Any sufficiently advanced technology is indistinguishable from magic"
\n\nClarke's 3'rd law
", "votes": 16, "replies": [] }, { "id": 139081, "author_name": "Giba", "author_url": "", "post_date": "2016-10-12T13:17:06.047000", "content": "Using only individual row information is hard to beat 0.31MCC and 0.78AUC and that model have a big value to Bosch. Using any kind of "data property feature" (there are many ways to explore it not using ID) it is possible to beat 0.5x, 0.93AUC. My current solution don't use IDs.
", "votes": 9, "replies": [] }, { "id": 138922, "author_name": "anokas", "author_url": "", "post_date": "2016-10-11T17:43:18.027000", "content": "[quote=gingerman;138897]
\n\nJust edited my answer at the first sentence was a fragment - @anokas reported not to be using ID at all and it seems plausible to me.
\n\n[/quote]
\n\nYep, you are correct. We had 0.921 AUC before we added our first ID/row order feature.
", "votes": 1, "replies": [] }, { "id": 138897, "author_name": "gingerman | Chris", "author_url": "", "post_date": "2016-10-11T14:57:02.530000", "content": "Just edited my answer at the first sentence was a fragment - @anokas reported not to be using ID at all and it seems plausible to me.
", "votes": 1, "replies": [] }, { "id": 138893, "author_name": "gingerman | Chris", "author_url": "", "post_date": "2016-10-11T14:31:37.117000", "content": "I guess this cannot be answered as one can indirectly leverage the leakage without using IDs. Still, @anokas has reported ~0.92 AUC and ~0.49 MCC without magic feature.
", "votes": 1, "replies": [] }, { "id": 139462, "author_name": "CPMP", "author_url": "", "post_date": "2016-10-14T09:09:26.617000", "content": "@Vicens,
\n\nyou just explained why the magic feature may have some meaningful, physical, reason to be there.
\n\nGreat insight.
", "votes": 2, "replies": [] }, { "id": 140728, "author_name": "CPMP", "author_url": "", "post_date": "2016-10-22T16:05:16.860000", "content": "@Jesse I am using 320 features only, including my engineered ones. Maybe I should use more ;)
\n\nIt would be great if an equivalent of Lasso is derived for GBM.
", "votes": 0, "replies": [] }, { "id": 140696, "author_name": "CPMP", "author_url": "", "post_date": "2016-10-22T10:34:44.333000", "content": "Jesse, there are various ways to capture that magic feature. Faron has shared one way, look for his kernel.
", "votes": 0, "replies": [] }, { "id": 139747, "author_name": "Sasho Nedelkoski", "author_url": "", "post_date": "2016-10-16T09:05:35.343000", "content": "Yes, there is higher probability but, clearly for now i don't see pattern for predicting the first, but there are patterns for second, third duplicates in a row. I am clearly missing something :)
", "votes": 0, "replies": [] }, { "id": 139672, "author_name": "Vicens Gaitan", "author_url": "", "post_date": "2016-10-15T18:49:42.617000", "content": "@Salka
\nyes, specially those with consecutive Id's. And you can bet for the first in the row ;)
@Vicens Gaitan
\n\n\n\n\nUsing time correlations ( but no row order correlations or ID's)
\n
By this you mean when duplicates ocurr there is greater probability of some product to fail ?
", "votes": 0, "replies": [] }, { "id": 139594, "author_name": "raman", "author_url": "", "post_date": "2016-10-15T03:56:31.850000", "content": "@Vicens, indeed thanks, very nice observations!
", "votes": 0, "replies": [] }, { "id": 139076, "author_name": "toshi_k", "author_url": "", "post_date": "2016-10-12T12:11:32.397000", "content": "Hi raman,
\n\nIn this competition, there are some types of magical features except for ID.
\nI barely reach 0.30 without magical features.
Thanks! How about without using the ID feature at all? I am trying to see how generalizable the data models can be. That is, can we predict for a random object only using the non-id data.
", "votes": 0, "replies": [] }, { "id": 140707, "author_name": "", "author_url": "", "post_date": "2016-10-22T12:19:57.327000", "content": "", "votes": 1, "replies": [] }, { "id": 140698, "author_name": "", "author_url": "", "post_date": "2016-10-22T10:37:59.387000", "content": "", "votes": 0, "replies": [] }, { "id": 140658, "author_name": "", "author_url": "", "post_date": "2016-10-22T01:42:20.773000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "139100": "Using time correlations ( but no row order correlations or ID's) and some *model based* feature engineering is possible to push the performance to **0.0281 logLoss / 0.807 AUC/ 0.317 MCC (0.3166 LB)**\r\n\r\nAnyway, some \"real\" phenomena can cause row duplicates in the data: after a comunication crash, SCADA ( the subsystem responsible for reading and storing the measurements) usually repeat the last seen value, so the measurements associated to a given sequence of part numbers corresponds to the last correctly received, until the comunications is recovered.\r\n\r\nSupose the comunication failure is produced by a power glitch ( or blackout) and this glitch also affect the parts quality currently manufactured. Then the first part number in a sequence of duplications will fail the quality check.\r\n\r\nDoesn't sounds familiar to you? ;)\r\n\r\nJust writing some science-fiction about \"magic data properties\"\r\n\r\n*\"Any sufficiently advanced technology is indistinguishable from magic\"*\r\n\r\n*Clarke's 3'rd law*", "139081": "Using only individual row information is hard to beat 0.31MCC and 0.78AUC and that model have a big value to Bosch. Using any kind of \"data property feature\" (there are many ways to explore it not using ID) it is possible to beat 0.5x, 0.93AUC. My current solution don't use IDs.", "138922": "[quote=gingerman;138897]\r\n\r\nJust edited my answer at the first sentence was a fragment - @anokas reported not to be using ID at all and it seems plausible to me.\r\n\r\n[/quote]\r\n\r\nYep, you are correct. We had 0.921 AUC before we added our first ID/row order feature.", "138897": "Just edited my answer at the first sentence was a fragment - @anokas reported not to be using ID at all and it seems plausible to me.", "138893": "I guess this cannot be answered as one can indirectly leverage the leakage without using IDs. Still, @anokas has reported ~0.92 AUC and ~0.49 MCC without magic feature.", "139462": "@Vicens,\r\n\r\nyou just explained why the magic feature may have some meaningful, physical, reason to be there.\r\n\r\nGreat insight.", "140728": "@Jesse I am using 320 features only, including my engineered ones. Maybe I should use more ;)\r\n\r\nIt would be great if an equivalent of Lasso is derived for GBM. \r\n", "140696": "Jesse, there are various ways to capture that magic feature. Faron has shared one way, look for his kernel.", "139747": "Yes, there is higher probability but, clearly for now i don't see pattern for predicting the first, but there are patterns for second, third duplicates in a row. I am clearly missing something :)", "139672": "@Salka \r\nyes, specially those with consecutive Id's. And you can bet for the first in the row ;) ", "139651": "@Vicens Gaitan\r\n\r\n> Using time correlations ( but no row order correlations or ID's)\r\n\r\nBy this you mean when duplicates ocurr there is greater probability of some product to fail ?", "139594": "@Vicens, indeed thanks, very nice observations!", "139076": "Hi raman,\r\n\r\nIn this competition, there are some types of magical features except for ID.