![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1048539%2F04e5de4a4a9a337f5c34f444fb5fcd8b%2FCapture.PNG?generation=1727469823753177&alt=media\")
As PCIAT variables data is missing from test set what should we do?
\ndo we remove all PCIAT from train as well ?
As PCIAT variables data is missing from test set what should we do?
\ndo we remove all PCIAT from train as well ?
This is probably the way most of us do itðŸ˜
", "rawMarkdown": "This is probably the way most of us do itðŸ˜", "votes": null }, { "id": "3000246", "postDate": "09/27/2024 12:39:04", "content": "My current belief is that test set true sii is based on PCIAT_total, but of course PCIAT_total in test set is hidden before disclosed to us.
\nI'm now trying to generalize PCIAT_total to all train data, and then using the train data to predict a PCIAT_total in test set, and convert it to sii for submission. Basically a LGBM on top of another LGBM.
\nOf course the flow is tedious, and the result is not promising yet.
\nGood luck!
This is the best option if you don't wish to structure sii from PCIAT-Total @chetan8007
", "rawMarkdown": "This is the best option if you don't wish to structure **sii** from **PCIAT-Total** @chetan8007", "votes": null }, { "id": "3000472", "postDate": "09/27/2024 16:48:50", "content": "Hi,
\n'sii' comes from PCIAT-total and PCIAT-total is the sum of the 20 PCIAT-PCiAT-xx.
\nInstead of fitting directly a model which predict 'sii', we can try to fit models to predict each of the 20 PCIAT-PCIAT, then compute PCIAT-total as the sum of PCIAT-PCIAT's predictions and then compute 'sii' predictions.
\nBecause there is several ways to have a high PCIAT-total value (or a high \"sii' value)
\ndisplay((train[] - train[[ i ()]].(axis = )).value_counts(dropna=))\n\n_ = {i: i ()}\n_.update({i: i (, )})\n_.update({i: i (, )})\n_.update({i: i (, )})\n\ndisplay((train[].(_) - train[]).value_counts(dropna=))\n \n \n \n\n \n \n \n@adaubas I tried this to no good effect. LB score is quite tepid.
", "rawMarkdown": "adaubas I tried this to no good effect. LB score is quite tepid.", "votes": null }, { "id": "3000620", "postDate": "09/27/2024 20:45:06", "content": "PCIAT-Season is NaN if and only if sii is NaN.
np.(train[].isna() ^ train[].isna())\n0\nHowever, even if PCIAT-Season is one of the valid values (Spring, Fall, Summer, Winter), some of the summands can be NaN, and PCIAT-PCIAT_Total is the nansum of them. If PCIAT-Season is NaN, then all the summands would be NaN, but PCIAT-PCIAT_Total would be NaN too, not the nansum (which would be 0).
So, when modeling as a multi-output regression/classification problem, one possible approach is, first remove all the rows with NaN PCIAT-Season, and then fillna all the summand columns in the remaining rows with 0 so that the multi-output target would not have any missing values. This is assuming that in the (hidden) test set, the latent PCIAT-Season variable would never be NaN, since sii is not supposed to be NaN in the test set.
X = train[~train[].isna()]\ndf = {}\ncols = [ i ()]\n col cols:\n df.update({col:X[col].value_counts(dropna=)})\npd.DataFrame(df)\n
Thank you for feedback @ravi20076
", "rawMarkdown": "Thank you for feedback @ravi20076", "votes": null }, { "id": "3000928", "postDate": "09/28/2024 09:14:20", "content": "Wow, nice catch @siukeitin ! That might explain some of the discrepancies between SII values and other features, I will check (UPD: i did, 17 rows have incorrect SII, but nothing explained, discrepancies persist)! It would be really helpful to know the reason why some questions of the Parent-Child Internet Addiction Test can be ignored…
", "rawMarkdown": "Wow, nice catch @siukeitin ! That might explain some of the discrepancies between SII values and other features, I will check (*UPD: i did, 17 rows have incorrect SII, but nothing explained, discrepancies persist*)! It would be really helpful to know the reason why some questions of the Parent-Child Internet Addiction Test can be ignored...", "votes": null }, { "id": "3001407", "postDate": "09/28/2024 20:00:58", "content": "Would the best option not be to use the PCIAT-Total score as the target as it can provide more information to the model about what scores are on the borders of the SII classes?
", "rawMarkdown": "Would the best option not be to use the PCIAT-Total score as the target as it can provide more information to the model about what scores are on the borders of the SII classes?", "votes": null }, { "id": "3002100", "postDate": "09/29/2024 15:58:01", "content": "As \"sii\" we have many missing values, so colud be the possiblity to take PICAT columns average value as a target value.. what you people think..
", "rawMarkdown": "As \"sii\" we have many missing values, so colud be the possiblity to take PICAT columns average value as a target value.. what you people think..", "votes": null }, { "id": "3002413", "postDate": "09/30/2024 02:39:58", "content": "No that won’t be possible, as usually both the target and PCIAT variables are missing.
", "rawMarkdown": "No that won’t be possible, as usually both the target and PCIAT variables are missing.", "votes": null }, { "id": "3003402", "postDate": "10/01/2024 01:24:48", "content": "Hi Antonina, can you share which discrepancy you found? I may be missing smth but find the sii corresponds to the sum converted. thanks!
", "rawMarkdown": "Hi Antonina, can you share which discrepancy you found? I may be missing smth but find the sii corresponds to the sum converted. thanks!", "votes": null }, { "id": "3003634", "postDate": "10/01/2024 06:16:51", "content": "Sorry if I didn't get your point, but just in case:
\nPCIAT-PCIAT_Total is calculated as the sum of the scores for the answers to 20 questions (stored in PCIAT-PCIAT_01 to PCIAT-PCIAT_20). Some of the questions may be ignored by a respondent (missing values in the PCIAT columns), but the SII score is still calculated as the sum of the non-NA values, resulting in potentially invalid SII values:
\ne.g. PCIAT-PCIAT_Total between 31 and 49 corresponds to SII = 1 and between 50 and 79 corresponds to SII = 2. A respondent has 47 points and did not answer 1 question. If he did, his PCIAT-PCIAT_Total could be 47 to 52 (max 5 points per question), so his actual SII could be 1 or 2. But in the data we have 1.
", "rawMarkdown": "Sorry if I didn't get your point, but just in case:\n\nPCIAT-PCIAT_Total is calculated as the sum of the scores for the answers to 20 questions (stored in PCIAT-PCIAT_01 to PCIAT-PCIAT_20). Some of the questions may be ignored by a respondent (missing values in the PCIAT columns), but the SII score is still calculated as the sum of the non-NA values, resulting in potentially invalid SII values:\n\ne.g. PCIAT-PCIAT_Total between 31 and 49 corresponds to SII = 1 and between 50 and 79 corresponds to SII = 2. A respondent has 47 points and did not answer 1 question. If he did, his PCIAT-PCIAT_Total could be 47 to 52 (max 5 points per question), so his actual SII could be 1 or 2. But in the data we have 1.", "votes": null }, { "id": "3006687", "postDate": "10/04/2024 12:28:58", "content": "@chetan8007 I've reviewed all the comments and haven't found any additional insights. If you've drawn any conclusions or developed a method for handling the PCIAT data or imputing the sii values, I would appreciate your guidance.
", "rawMarkdown": "chetan8007 I've reviewed all the comments and haven't found any additional insights. If you've drawn any conclusions or developed a method for handling the PCIAT data or imputing the sii values, I would appreciate your guidance.", "votes": null }, { "id": "3008224", "postDate": "10/06/2024 11:35:01", "content": "Hello,I'm doing this job ,and I want to know what't your LB score by predicting every PCIAT?
", "rawMarkdown": "Hello,I'm doing this job ,and I want to know what't your LB score by predicting every PCIAT?", "votes": null }, { "id": "3073986", "postDate": "12/17/2024 07:24:22", "content": "Thank you for your reply
", "rawMarkdown": "Thank you for your reply", "votes": null } ], "comments": [ { "id": 3000226, "author_name": "wayne127", "author_url": "", "post_date": "09/27/2024 12:16:23", "content": "This is probably the way most of us do itðŸ˜
", "votes": null, "replies": [ { "id": 3002100, "author_name": "sadiasadique", "author_url": "", "post_date": "09/29/2024 15:58:01", "content": "As \"sii\" we have many missing values, so colud be the possiblity to take PICAT columns average value as a target value.. what you people think..
", "votes": null, "replies": [ { "id": 3002413, "author_name": "peterhopkinson", "author_url": "", "post_date": "09/30/2024 02:39:58", "content": "No that won’t be possible, as usually both the target and PCIAT variables are missing.
", "votes": null, "replies": [] } ] } ] }, { "id": 3000246, "author_name": "tomyuen", "author_url": "", "post_date": "09/27/2024 12:39:04", "content": "My current belief is that test set true sii is based on PCIAT_total, but of course PCIAT_total in test set is hidden before disclosed to us.
\nI'm now trying to generalize PCIAT_total to all train data, and then using the train data to predict a PCIAT_total in test set, and convert it to sii for submission. Basically a LGBM on top of another LGBM.
\nOf course the flow is tedious, and the result is not promising yet.
\nGood luck!
This is the best option if you don't wish to structure sii from PCIAT-Total @chetan8007
", "votes": null, "replies": [ { "id": 3001407, "author_name": "peterhopkinson", "author_url": "", "post_date": "09/28/2024 20:00:58", "content": "Would the best option not be to use the PCIAT-Total score as the target as it can provide more information to the model about what scores are on the borders of the SII classes?
", "votes": null, "replies": [] } ] }, { "id": 3000472, "author_name": "adaubas", "author_url": "", "post_date": "09/27/2024 16:48:50", "content": "Hi,
\n'sii' comes from PCIAT-total and PCIAT-total is the sum of the 20 PCIAT-PCiAT-xx.
\nInstead of fitting directly a model which predict 'sii', we can try to fit models to predict each of the 20 PCIAT-PCIAT, then compute PCIAT-total as the sum of PCIAT-PCIAT's predictions and then compute 'sii' predictions.
\nBecause there is several ways to have a high PCIAT-total value (or a high \"sii' value)
\ndisplay((train[] - train[[ i ()]].(axis = )).value_counts(dropna=))\n\n_ = {i: i ()}\n_.update({i: i (, )})\n_.update({i: i (, )})\n_.update({i: i (, )})\n\ndisplay((train[].(_) - train[]).value_counts(dropna=))\n \n \n \n\n \n \n \n@adaubas I tried this to no good effect. LB score is quite tepid.
", "votes": null, "replies": [ { "id": 3000864, "author_name": "adaubas", "author_url": "", "post_date": "09/28/2024 06:55:24", "content": "Thank you for feedback @ravi20076
", "votes": null, "replies": [] }, { "id": 3008224, "author_name": "yashi003", "author_url": "", "post_date": "10/06/2024 11:35:01", "content": "Hello,I'm doing this job ,and I want to know what't your LB score by predicting every PCIAT?
", "votes": null, "replies": [] } ] }, { "id": 3000620, "author_name": "siukeitin", "author_url": "", "post_date": "09/27/2024 20:45:06", "content": "PCIAT-Season is NaN if and only if sii is NaN.
np.(train[].isna() ^ train[].isna())\n0\nHowever, even if PCIAT-Season is one of the valid values (Spring, Fall, Summer, Winter), some of the summands can be NaN, and PCIAT-PCIAT_Total is the nansum of them. If PCIAT-Season is NaN, then all the summands would be NaN, but PCIAT-PCIAT_Total would be NaN too, not the nansum (which would be 0).
So, when modeling as a multi-output regression/classification problem, one possible approach is, first remove all the rows with NaN PCIAT-Season, and then fillna all the summand columns in the remaining rows with 0 so that the multi-output target would not have any missing values. This is assuming that in the (hidden) test set, the latent PCIAT-Season variable would never be NaN, since sii is not supposed to be NaN in the test set.
X = train[~train[].isna()]\ndf = {}\ncols = [ i ()]\n col cols:\n df.update({col:X[col].value_counts(dropna=)})\npd.DataFrame(df)\n
Wow, nice catch @siukeitin ! That might explain some of the discrepancies between SII values and other features, I will check (UPD: i did, 17 rows have incorrect SII, but nothing explained, discrepancies persist)! It would be really helpful to know the reason why some questions of the Parent-Child Internet Addiction Test can be ignored…
", "votes": null, "replies": [ { "id": 3003402, "author_name": "ibabic", "author_url": "", "post_date": "10/01/2024 01:24:48", "content": "Hi Antonina, can you share which discrepancy you found? I may be missing smth but find the sii corresponds to the sum converted. thanks!
", "votes": null, "replies": [ { "id": 3003634, "author_name": "antoninadolgorukova", "author_url": "", "post_date": "10/01/2024 06:16:51", "content": "Sorry if I didn't get your point, but just in case:
\nPCIAT-PCIAT_Total is calculated as the sum of the scores for the answers to 20 questions (stored in PCIAT-PCIAT_01 to PCIAT-PCIAT_20). Some of the questions may be ignored by a respondent (missing values in the PCIAT columns), but the SII score is still calculated as the sum of the non-NA values, resulting in potentially invalid SII values:
\ne.g. PCIAT-PCIAT_Total between 31 and 49 corresponds to SII = 1 and between 50 and 79 corresponds to SII = 2. A respondent has 47 points and did not answer 1 question. If he did, his PCIAT-PCIAT_Total could be 47 to 52 (max 5 points per question), so his actual SII could be 1 or 2. But in the data we have 1.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3073986, "author_name": "prosepoem", "author_url": "", "post_date": "12/17/2024 07:24:22", "content": "Thank you for your reply
", "votes": null, "replies": [] } ] }, { "id": 3006687, "author_name": "nassimsfaxi", "author_url": "", "post_date": "10/04/2024 12:28:58", "content": "@chetan8007 I've reviewed all the comments and haven't found any additional insights. If you've drawn any conclusions or developed a method for handling the PCIAT data or imputing the sii values, I would appreciate your guidance.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3000125": "As PCIAT variables data is missing from test set what should we do?\ndo we remove all PCIAT from train as well ?", "3000226": "This is probably the way most of us do itðŸ˜", "3000246": "My current belief is that test set true sii is based on PCIAT_total, but of course PCIAT_total in test set is hidden before disclosed to us.\nI'm now trying to generalize PCIAT_total to all train data, and then using the train data to predict a PCIAT_total in test set, and convert it to sii for submission. Basically a LGBM on top of another LGBM.\nOf course the flow is tedious, and the result is not promising yet.\nGood luck!", "3000248": "This is the best option if you don't wish to structure **sii** from **PCIAT-Total** @chetan8007", "3000472": "Hi,\n\n'sii' comes from PCIAT-total and PCIAT-total is the sum of the 20 PCIAT-PCiAT-xx.\n\nInstead of fitting directly a model which predict 'sii', we can try to fit models to predict each of the 20 PCIAT-PCIAT, then compute PCIAT-total as the sum of PCIAT-PCIAT's predictions and then compute 'sii' predictions.\n\nBecause there is several ways to have a high PCIAT-total value (or a high \"sii' value)\n\n```python\ndisplay((train[\"PCIAT-PCIAT_Total\"] - train[[f\"PCIAT-PCIAT_{i+1:02}\" for i in range(20)]].sum(axis = 1)).value_counts(dropna=False))\n\n_map = {i:0 for i in range(31)}\n_map.update({i:1 for i in range(31, 50)})\n_map.update({i:2 for i in range(50, 80)})\n_map.update({i:3 for i in range(80, 101)})\n\ndisplay((train[\"PCIAT-PCIAT_Total\"].map(_map) - train['sii']).value_counts(dropna=False))\n```\n```\n0.0 2736\nNaN 1224\nName: count, dtype: int64\n\n0.0 2736\nNaN 1224\nName: count, dtype: int64\n```", "3000602": "adaubas I tried this to no good effect. LB score is quite tepid.", "3000620": "`PCIAT-Season` is NaN if and only if `sii` is NaN. \n```python\nnp.sum(train['PCIAT-Season'].isna() ^ train['sii'].isna())\n```\n```text\n0\n```\nHowever, even if `PCIAT-Season` is one of the valid values (`Spring`, `Fall`, `Summer`, `Winter`), some of the summands can be NaN, and `PCIAT-PCIAT_Total` is the `nansum` of them. If `PCIAT-Season` is NaN, then all the summands would be NaN, but `PCIAT-PCIAT_Total` would be NaN too, not the `nansum` (which would be 0).\n\nSo, when modeling as a multi-output regression/classification problem, one possible approach is, first remove all the rows with NaN `PCIAT-Season`, and then `fillna` all the summand columns in the remaining rows with 0 so that the multi-output target would not have any missing values. This is assuming that in the (hidden) test set, the latent `PCIAT-Season` variable would never be NaN, since `sii` is not supposed to be NaN in the test set.\n```python\nX = train[~train['PCIAT-Season'].isna()]\ndf = {}\ncols = [F'PCIAT-PCIAT_{i+1:02d}' for i in range(20)]\nfor col in cols:\n df.update({col:X[col].value_counts(dropna=False)})\npd.DataFrame(df)\n```\n", "3000864": "Thank you for feedback @ravi20076", "3000928": "Wow, nice catch @siukeitin ! That might explain some of the discrepancies between SII values and other features, I will check (*UPD: i did, 17 rows have incorrect SII, but nothing explained, discrepancies persist*)! It would be really helpful to know the reason why some questions of the Parent-Child Internet Addiction Test can be ignored...", "3001407": "Would the best option not be to use the PCIAT-Total score as the target as it can provide more information to the model about what scores are on the borders of the SII classes?", "3002100": "As \"sii\" we have many missing values, so colud be the possiblity to take PICAT columns average value as a target value.. what you people think..", "3002413": "No that won’t be possible, as usually both the target and PCIAT variables are missing.", "3003402": "Hi Antonina, can you share which discrepancy you found? I may be missing smth but find the sii corresponds to the sum converted. thanks!", "3003634": "Sorry if I didn't get your point, but just in case:\n\nPCIAT-PCIAT_Total is calculated as the sum of the scores for the answers to 20 questions (stored in PCIAT-PCIAT_01 to PCIAT-PCIAT_20). Some of the questions may be ignored by a respondent (missing values in the PCIAT columns), but the SII score is still calculated as the sum of the non-NA values, resulting in potentially invalid SII values:\n\ne.g. PCIAT-PCIAT_Total between 31 and 49 corresponds to SII = 1 and between 50 and 79 corresponds to SII = 2. A respondent has 47 points and did not answer 1 question. If he did, his PCIAT-PCIAT_Total could be 47 to 52 (max 5 points per question), so his actual SII could be 1 or 2. But in the data we have 1.", "3006687": "chetan8007 I've reviewed all the comments and haven't found any additional insights. If you've drawn any conclusions or developed a method for handling the PCIAT data or imputing the sii values, I would appreciate your guidance.", "3008224": "Hello,I'm doing this job ,and I want to know what't your LB score by predicting every PCIAT?", "3073986": "Thank you for your reply" }, "source": "meta" }