I found that -33.21928095 comes from -10/lg2, where 2^33.21928095=10,00,000,000 or in other words, 2^-33.21928095=10^-10 (very close to 0). So I am thinking how is the loyalty score calculated? Does -33.21928095 mean customer is not loyal at all? But I found card_id='C_ID_e34ea29439', whose target value is -33.21928095, had 136 records in historical_transactions and 2 records in historical_transactions
", "messages": [ { "id": "462199", "postDate": "01/27/2019 21:06:30", "content": "I found that -33.21928095 comes from -10/lg2, where 2^33.21928095=10,00,000,000 or in other words, 2^-33.21928095=10^-10 (very close to 0). So I am thinking how is the loyalty score calculated? Does -33.21928095 mean customer is not loyal at all? But I found card_id='C_ID_e34ea29439', whose target value is -33.21928095, had 136 records in historical_transactions and 2 records in historical_transactions
", "rawMarkdown": "I found that -33.21928095 comes from -10/lg2, where 2^33.21928095=10,00,000,000 or in other words, 2^-33.21928095=10^-10 (very close to 0). So I am thinking how is the loyalty score calculated? Does -33.21928095 mean customer is not loyal at all? But I found card_id='C_ID_e34ea29439', whose target value is -33.21928095, had 136 records in historical_transactions and 2 records in historical_transactions", "votes": null }, { "id": "462207", "postDate": "01/27/2019 21:34:02", "content": "-33.21928095 could be nan equivalent or -10 in some scale or sum of answers from questionnaire in scale fro 0 to 10 or sum of points in cardholder account or or or
\n\nWhy NaN\n 1. Check historical transactions\n 2. Remove Cards that are in new_merchant\n 3. Agg by card_id (size)\n 4. Check cards with size 2,3,4...\nThere are some similar profiles but some have -33 and some positive loyalty score (with some variation there should be 1-2-5 points difference but not 30)
\n\nAnd all other theories have some base but we dont know what is Elo Score and how it is calculated.
\n\n~~~~\nstrange_items = historical_df[~(historical_df['card_id'].isin(new_merchant_df['card_id']))]\\\n .groupby(['card_id'])['authorized_flag'].agg(['size'])\\\n .reset_index().rename(columns={'size': 'number_of_transactions'})\nstrange_items = strange_items[strange_items['number_of_transactions']<7]\nstrange_items = train[(train['card_id'].isin(strange_items['card_id']))&((train['target']<-33)|(train['target']>0))]\n~~~~
", "rawMarkdown": "33.21928095 could be nan equivalent or -10 in some scale or sum of answers from questionnaire in scale fro 0 to 10 or sum of points in cardholder account or or or \n\nWhy NaN\n 1. Check historical transactions\n 2. Remove Cards that are in new_merchant\n 3. Agg by card_id (size)\n 4. Check cards with size 2,3,4...\nThere are some similar profiles but some have -33 and some positive loyalty score (with some variation there should be 1-2-5 points difference but not 30)\n\nAnd all other theories have some base but we dont know what is Elo Score and how it is calculated.\n\n~~~~\nstrange_items = historical_df[~(historical_df['card_id'].isin(new_merchant_df['card_id']))]\\\n .groupby(['card_id'])['authorized_flag'].agg(['size'])\\\n .reset_index().rename(columns={'size': 'number_of_transactions'})\nstrange_items = strange_items[strange_items['number_of_transactions']<7]\nstrange_items = train[(train['card_id'].isin(strange_items['card_id']))&((train['target']<-33)|(train['target']>0))]\n~~~~", "votes": null }, { "id": "462783", "postDate": "01/28/2019 22:10:09", "content": "I tried to train the model without those outliers, replaced by mean, median. The score on train and cv reduced to 1.5~1.6, but LB went to 3.8+
", "rawMarkdown": "I tried to train the model without those outliers, replaced by mean, median. The score on train and cv reduced to 1.5~1.6, but LB went to 3.8+", "votes": null }, { "id": "462826", "postDate": "01/29/2019 01:09:58", "content": "This is quite difficult. I am still confused about this.
", "rawMarkdown": "This is quite difficult. I am still confused about this.", "votes": null }, { "id": "463375", "postDate": "01/29/2019 22:15:44", "content": "@jieWu Yeah, Removing the outliers and training the model doesn't seem a good practice here. Though you can use this method for post processing of predictions. That might help you.
", "rawMarkdown": "jieWu Yeah, Removing the outliers and training the model doesn't seem a good practice here. Though you can use this method for post processing of predictions. That might help you.", "votes": null }, { "id": "463387", "postDate": "01/29/2019 22:46:47", "content": "If outliers were NaNs, there would most likely be a single value for all outliers, and probably, it would be an integer, not a float.\nAlso, if you look at a histogram of values below -33, the shape of the histogram is very similar to the shape of histogram for the non-outliers. It might mean that these values carry some sort of a signal, and the signal is important. This would explain (to me at least) why removing the outliers impacts the LB score so much
", "rawMarkdown": "If outliers were NaNs, there would most likely be a single value for all outliers, and probably, it would be an integer, not a float.\nAlso, if you look at a histogram of values below -33, the shape of the histogram is very similar to the shape of histogram for the non-outliers. It might mean that these values carry some sort of a signal, and the signal is important. This would explain (to me at least) why removing the outliers impacts the LB score so much", "votes": null }, { "id": "463403", "postDate": "01/29/2019 23:44:43", "content": "Follow idea from https://www.kaggle.com/waitingli/combining-your-model-with-a-model-without-outlier I did these:\n1: I tried to build models either without 'outliers' or replace 'outliers' by mean, and use those models to predict test data. 2: Build a classification model to predict 'outlier'. 3: (My best model) Build a model with 'outlier', and predict on test data. 4: Use target value from 3 to replace the target value from 1, if the card_id is predicted (from 2) as 'outlier'. But this did not improve my best result at all.
", "rawMarkdown": "Follow idea from https://www.kaggle.com/waitingli/combining-your-model-with-a-model-without-outlier I did these:\n1: I tried to build models either without 'outliers' or replace 'outliers' by mean, and use those models to predict test data. 2: Build a classification model to predict 'outlier'. 3: (My best model) Build a model with 'outlier', and predict on test data. 4: Use target value from 3 to replace the target value from 1, if the card_id is predicted (from 2) as 'outlier'. But this did not improve my best result at all.", "votes": null }, { "id": "463404", "postDate": "01/29/2019 23:45:51", "content": "Yes, those so called 'outlier' seems are very important.
", "rawMarkdown": "Yes, those so called 'outlier' seems are very important.", "votes": null }, { "id": "465973", "postDate": "02/04/2019 12:55:01", "content": "I am pretty sure the -33 comes from someone log(x+epsilon) the response instead of log(x+1) where epsilonis a small number. In theat case -33 arrises if x=0.
", "rawMarkdown": "I am pretty sure the -33 comes from someone log(x+epsilon) the response instead of log(x+1) where epsilonis a small number. In theat case -33 arrises if x=0.", "votes": null }, { "id": "466084", "postDate": "02/04/2019 16:43:12", "content": "Yes, it's log(10^-10)
", "rawMarkdown": "Yes, it's log(10^-10)", "votes": null } ], "comments": [ { "id": 462207, "author_name": "kyakovlev", "author_url": "", "post_date": "01/27/2019 21:34:02", "content": "-33.21928095 could be nan equivalent or -10 in some scale or sum of answers from questionnaire in scale fro 0 to 10 or sum of points in cardholder account or or or
\n\nWhy NaN\n 1. Check historical transactions\n 2. Remove Cards that are in new_merchant\n 3. Agg by card_id (size)\n 4. Check cards with size 2,3,4...\nThere are some similar profiles but some have -33 and some positive loyalty score (with some variation there should be 1-2-5 points difference but not 30)
\n\nAnd all other theories have some base but we dont know what is Elo Score and how it is calculated.
\n\n~~~~\nstrange_items = historical_df[~(historical_df['card_id'].isin(new_merchant_df['card_id']))]\\\n .groupby(['card_id'])['authorized_flag'].agg(['size'])\\\n .reset_index().rename(columns={'size': 'number_of_transactions'})\nstrange_items = strange_items[strange_items['number_of_transactions']<7]\nstrange_items = train[(train['card_id'].isin(strange_items['card_id']))&((train['target']<-33)|(train['target']>0))]\n~~~~
", "votes": null, "replies": [ { "id": 462783, "author_name": "lftuwujie", "author_url": "", "post_date": "01/28/2019 22:10:09", "content": "I tried to train the model without those outliers, replaced by mean, median. The score on train and cv reduced to 1.5~1.6, but LB went to 3.8+
", "votes": null, "replies": [] }, { "id": 463375, "author_name": "harshit92", "author_url": "", "post_date": "01/29/2019 22:15:44", "content": "@jieWu Yeah, Removing the outliers and training the model doesn't seem a good practice here. Though you can use this method for post processing of predictions. That might help you.
", "votes": null, "replies": [] }, { "id": 463404, "author_name": "lftuwujie", "author_url": "", "post_date": "01/29/2019 23:45:51", "content": "Yes, those so called 'outlier' seems are very important.
", "votes": null, "replies": [] } ] }, { "id": 462826, "author_name": "longyin2", "author_url": "", "post_date": "01/29/2019 01:09:58", "content": "This is quite difficult. I am still confused about this.
", "votes": null, "replies": [] }, { "id": 463387, "author_name": "calabist", "author_url": "", "post_date": "01/29/2019 22:46:47", "content": "If outliers were NaNs, there would most likely be a single value for all outliers, and probably, it would be an integer, not a float.\nAlso, if you look at a histogram of values below -33, the shape of the histogram is very similar to the shape of histogram for the non-outliers. It might mean that these values carry some sort of a signal, and the signal is important. This would explain (to me at least) why removing the outliers impacts the LB score so much
", "votes": null, "replies": [ { "id": 463403, "author_name": "lftuwujie", "author_url": "", "post_date": "01/29/2019 23:44:43", "content": "Follow idea from https://www.kaggle.com/waitingli/combining-your-model-with-a-model-without-outlier I did these:\n1: I tried to build models either without 'outliers' or replace 'outliers' by mean, and use those models to predict test data. 2: Build a classification model to predict 'outlier'. 3: (My best model) Build a model with 'outlier', and predict on test data. 4: Use target value from 3 to replace the target value from 1, if the card_id is predicted (from 2) as 'outlier'. But this did not improve my best result at all.
", "votes": null, "replies": [] } ] }, { "id": 465973, "author_name": "trainedonatoaster", "author_url": "", "post_date": "02/04/2019 12:55:01", "content": "I am pretty sure the -33 comes from someone log(x+epsilon) the response instead of log(x+1) where epsilonis a small number. In theat case -33 arrises if x=0.
", "votes": null, "replies": [ { "id": 466084, "author_name": "lftuwujie", "author_url": "", "post_date": "02/04/2019 16:43:12", "content": "Yes, it's log(10^-10)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "462199": "I found that -33.21928095 comes from -10/lg2, where 2^33.21928095=10,00,000,000 or in other words, 2^-33.21928095=10^-10 (very close to 0). So I am thinking how is the loyalty score calculated? Does -33.21928095 mean customer is not loyal at all? But I found card_id='C_ID_e34ea29439', whose target value is -33.21928095, had 136 records in historical_transactions and 2 records in historical_transactions", "462207": "33.21928095 could be nan equivalent or -10 in some scale or sum of answers from questionnaire in scale fro 0 to 10 or sum of points in cardholder account or or or \n\nWhy NaN\n 1. Check historical transactions\n 2. Remove Cards that are in new_merchant\n 3. Agg by card_id (size)\n 4. Check cards with size 2,3,4...\nThere are some similar profiles but some have -33 and some positive loyalty score (with some variation there should be 1-2-5 points difference but not 30)\n\nAnd all other theories have some base but we dont know what is Elo Score and how it is calculated.\n\n~~~~\nstrange_items = historical_df[~(historical_df['card_id'].isin(new_merchant_df['card_id']))]\\\n .groupby(['card_id'])['authorized_flag'].agg(['size'])\\\n .reset_index().rename(columns={'size': 'number_of_transactions'})\nstrange_items = strange_items[strange_items['number_of_transactions']<7]\nstrange_items = train[(train['card_id'].isin(strange_items['card_id']))&((train['target']<-33)|(train['target']>0))]\n~~~~", "462783": "I tried to train the model without those outliers, replaced by mean, median. The score on train and cv reduced to 1.5~1.6, but LB went to 3.8+", "462826": "This is quite difficult. I am still confused about this.", "463375": "jieWu Yeah, Removing the outliers and training the model doesn't seem a good practice here. Though you can use this method for post processing of predictions. That might help you.", "463387": "If outliers were NaNs, there would most likely be a single value for all outliers, and probably, it would be an integer, not a float.\nAlso, if you look at a histogram of values below -33, the shape of the histogram is very similar to the shape of histogram for the non-outliers. It might mean that these values carry some sort of a signal, and the signal is important. This would explain (to me at least) why removing the outliers impacts the LB score so much", "463403": "Follow idea from https://www.kaggle.com/waitingli/combining-your-model-with-a-model-without-outlier I did these:\n1: I tried to build models either without 'outliers' or replace 'outliers' by mean, and use those models to predict test data. 2: Build a classification model to predict 'outlier'. 3: (My best model) Build a model with 'outlier', and predict on test data. 4: Use target value from 3 to replace the target value from 1, if the card_id is predicted (from 2) as 'outlier'. But this did not improve my best result at all.", "463404": "Yes, those so called 'outlier' seems are very important.", "465973": "I am pretty sure the -33 comes from someone log(x+epsilon) the response instead of log(x+1) where epsilonis a small number. In theat case -33 arrises if x=0.", "466084": "Yes, it's log(10^-10)" }, "source": "meta" }