![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113389%2F8d592f77b0469f021312687c9dc76ad8%2FScreenshot%202020-10-31%20at%203.38.32%20PM.png?generation=1604139023264162&alt=media\")
What I have tested:
\nBTW: All these features improve score
", "messages": [ { "id": "1054945", "postDate": "10/20/2020 10:24:01", "content": "What I have tested:
\nBTW: All these features improve score
", "rawMarkdown": "What I have tested:\n\n1. Flags that show that user watched at least one lecture of each part and each type (11 features)\n2. Flag that shows that user watched at least one lecture\n3. Merge 'bundle_id' and 'part' from questions data set\n4. Fill 'prior_question_elapsed_time' with mean\n\nBTW: All these features improve score", "votes": null }, { "id": "1054980", "postDate": "10/20/2020 11:37:12", "content": "I'm trying
\nNot sure about the feature importance yet.
", "rawMarkdown": "I'm trying \n- mean bundle accuracy\n- mean part accuracy\n- tag, user_id, bundle_id and part embeddings\n\nNot sure about the feature importance yet.", "votes": null }, { "id": "1055027", "postDate": "10/20/2020 12:22:20", "content": "I am using Target Encoding for:
\nHow about seq2seq model
", "rawMarkdown": "How about seq2seq model", "votes": null }, { "id": "1055393", "postDate": "10/20/2020 18:12:38", "content": "\n\nFlag that shows that user watched at least one lecture
\n
You might also want to create one feature per lecture type:
\nI'm just trying to make features about congeniality between users and questions.
\nSpecifically, I compare the quesion's average correct answer rate for each \"part\" with the user's one.
\nI'm just trying now, so Not sure about the feature importance yet.
I have 11 features:
\n['part_1_boolean',\n 'part_2_boolean',\n 'part_3_boolean',\n 'part_4_boolean',\n 'part_5_boolean',\n 'part_6_boolean',\n 'part_7_boolean',\n 'type_of_concept_boolean',\n 'type_of_intention_boolean',\n 'type_of_solving_question_boolean',\n 'type_of_starter_boolean']\nfor lecture types and for parts.
", "rawMarkdown": "I have 11 features:\n\n```\n['part_1_boolean',\n 'part_2_boolean',\n 'part_3_boolean',\n 'part_4_boolean',\n 'part_5_boolean',\n 'part_6_boolean',\n 'part_7_boolean',\n 'type_of_concept_boolean',\n 'type_of_intention_boolean',\n 'type_of_solving_question_boolean',\n 'type_of_starter_boolean']\n```\n\nfor lecture types and for parts.", "votes": null }, { "id": "1056354", "postDate": "10/21/2020 16:13:26", "content": "I am trying to split the users to categories: smart ones, with high percentage of correctly answered questions and not. It is also correlating with summary time spent on the platform by the user.
", "rawMarkdown": "I am trying to split the users to categories: smart ones, with high percentage of correctly answered questions and not. It is also correlating with summary time spent on the platform by the user.", "votes": null }, { "id": "1056355", "postDate": "10/21/2020 16:14:37", "content": "Also I found the users which are both in train and test. I think, this could be useful.
", "rawMarkdown": "Also I found the users which are both in train and test. I think, this could be useful.", "votes": null }, { "id": "1056422", "postDate": "10/21/2020 17:26:52", "content": "@anaidashaginian Related to this, you might find this notebook interesting: https://www.kaggle.com/datafan07/riiid-challenge-eda-baseline-model#Answer-Accuracy---Time-Relations
", "rawMarkdown": "anaidashaginian Related to this, you might find this notebook interesting: https://www.kaggle.com/datafan07/riiid-challenge-eda-baseline-model#Answer-Accuracy---Time-Relations", "votes": null }, { "id": "1056427", "postDate": "10/21/2020 17:32:19", "content": "@anaidashaginian is that feature working for you? It's not working for me: https://www.kaggle.com/c/riiid-test-answer-prediction/discussion/190217
", "rawMarkdown": "anaidashaginian is that feature working for you? It's not working for me: https://www.kaggle.com/c/riiid-test-answer-prediction/discussion/190217", "votes": null }, { "id": "1056748", "postDate": "10/22/2020 03:35:50", "content": "@anaidashaginian Both in public and private test data??
", "rawMarkdown": "anaidashaginian Both in public and private test data??", "votes": null }, { "id": "1060628", "postDate": "10/26/2020 12:18:02", "content": "What about trying to figure out a sort of \"learning curves\"? Something like a cumulative number of lectures or questions related to each part or some group of tags (they also can be clustered somehow).
\nI'm working on it now.
Will look into it.
", "rawMarkdown": "Will look into it.", "votes": null }, { "id": "1063734", "postDate": "10/29/2020 08:57:11", "content": "Does anyone manage to extract relevant information from the user's answer ?
", "rawMarkdown": "Does anyone manage to extract relevant information from the user's answer ?", "votes": null }, { "id": "1063749", "postDate": "10/29/2020 09:11:00", "content": "because testcase donot contain user's answer
", "rawMarkdown": "because testcase donot contain user's answer", "votes": null }, { "id": "1063754", "postDate": "10/29/2020 09:19:36", "content": "Test data does contain user's answer in the prior_group_responses field.
I'm looking to update on the fly user's answer statistics in test set like some of us do for user answered correctly
", "rawMarkdown": "I'm looking to update on the fly user's answer statistics in test set like some of us do for user answered correctly", "votes": null }, { "id": "1063765", "postDate": "10/29/2020 09:41:25", "content": "I'm curious about the team name \"question the answers\" in LB btw
", "rawMarkdown": "I'm curious about the team name \"question the answers\" in LB btw", "votes": null }, { "id": "1063978", "postDate": "10/29/2020 14:58:05", "content": "\n\nI'm looking to update on the fly user's answer statistics in test set like some of us do for user answered correctly
\n
Had tried a simple decay on past feats but it didn't help as updating the old stats with new will have little effect on the old one unless you compute them over a good sample size is my current understanding .
", "rawMarkdown": ">I'm looking to update on the fly user's answer statistics in test set like some of us do for user answered correctly\n\nHad tried a simple decay on past feats but it didn't help as updating the old stats with new will have little effect on the old one unless you compute them over a good sample size is my current understanding .", "votes": null }, { "id": "1065125", "postDate": "10/30/2020 21:48:24", "content": "I've implemented an expanding window average of target variable by user. I wrapped it with numba ant it runs all 100m rows in a minute.
\nimport numba\n\n@numba.jit(nopython=True)\ndef expanding_target_mean(data):\n \"\"\"\n it takes user_id and answered correctly columns\n and calculates expanding mean over target\n \"\"\"\n res = np.zeros_like(data[:,0])\n user_id = 0\n counter = 0\n for i in range(len(res)):\n if data[i, 0] != user_id:\n user_id = data[i, 0]\n res[i] = data[i, 1]\n counter = 0\n continue\n if data[i, 0] == user_id:\n counter += 1\n res[i] = np.mean(data[i-counter:i, 1])\n return res\n\ndata[\"expanding_target_mean\"] = expanding_target_mean(data[['user_id', 'answered_correctly']].astype(float).values)\nWhat is the res variable?
I have observed that content_id and bundle_id are the same thing if I am not missing something. So bundle_id does not add much?
", "rawMarkdown": "I have observed that content_id and bundle_id are the same thing if I am not missing something. So bundle_id does not add much?", "votes": null }, { "id": "1065462", "postDate": "10/31/2020 10:10:34", "content": "They aren't always the same. Some examples:
\n
Thanks for the detailed table. It seems bundle id can take same value for different question ids but these different values are almost same as question id (<1%). Mathematically they are linearly correalated in the data, you can see it if you plot it. So not sure if it will add any additional info to the model.
", "rawMarkdown": "Thanks for the detailed table. It seems bundle id can take same value for different question ids but these different values are almost same as question id (<1%). Mathematically they are linearly correalated in the data, you can see it if you plot it. So not sure if it will add any additional info to the model.", "votes": null }, { "id": "1069157", "postDate": "11/04/2020 06:41:00", "content": "I feel t's crucial to engineer features which one can translate to the test set optimally as well..
", "rawMarkdown": "I feel t's crucial to engineer features which one can translate to the test set optimally as well..", "votes": null }, { "id": "1075040", "postDate": "11/11/2020 10:52:23", "content": "So I tried the following:
\nX[\"correct_user_sum\"] = X[[\"user_id\",'answered_correctly']].groupby([\"user_id\"])['answered_correctly'].cumsum()\nX['correct_user_sum'] = X['correct_user_sum']-1\nLast line is to remove leakage
\nThis feature performs really really well localy (around 0.80 AUC), but very very poorly on the LB, I have no idea why.
\nI've already tried different CV strategies and it always performs the same
How are you dealing with the new information and new users in the test set during submission?
", "rawMarkdown": "How are you dealing with the new information and new users in the test set during submission?", "votes": null }, { "id": "1075062", "postDate": "11/11/2020 11:11:36", "content": "Tried simple merge and FillNa, then tried just a simple [correct_user_sum+answered_correctly]. Both cases gave the same result. All my features are put into Series, so it's easy to update and merge
", "rawMarkdown": "Tried simple merge and FillNa, then tried just a simple [correct_user_sum+answered_correctly]. Both cases gave the same result. All my features are put into Series, so it's easy to update and merge", "votes": null }, { "id": "1075065", "postDate": "11/11/2020 11:14:41", "content": "this is my update funtion that gets a Df with User-ID in previous loop and there answers in the next loop
\ndef registrar(df):\n reg = cudf.DataFrame()\n reg['user_id'] = df['user_id']\n reg['answered_correctly'] = df['answered_correctly']\n reg['answered_wrong'] = 1-reg['answered_correctly']\n #reg = cudf.merge(reg, cudf.DataFrame(wrong_user_sum[[\"wrong_user_sum\",'user_id']]), right_on='user_id',left_on='user_id',how='left')\n #reg = cudf.merge(reg, cudf.DataFrame(corrrect_user_sum[[\"correct_user_sum\",'user_id']]), right_on='user_id',left_on='user_id',how='left')\n reg = cudf.merge(reg, cudf.DataFrame(answer_wrong_user[['user_id','answer_wrong_user']]), right_on='user_id',left_on='user_id',how='left')\n reg = cudf.merge(reg, cudf.DataFrame(correct_user[['user_id','answered_correctly_user']]), right_on='user_id',left_on='user_id',how='left')\n reg = cudf.merge(reg, cudf.DataFrame(user_questions[['user_id','User_questions']]), right_on='user_id',left_on='user_id',how='left')\n\n reg['User_questions'].fillna(0,inplace=True)\n reg['answered_correctly_user'].fillna(np.float32(0.643215),inplace=True)\n reg['answer_wrong_user'].fillna(np.float32(1-0.643215),inplace=True)\n #reg['wrong_user_sum'].fillna(0,inplace=True)\n #reg['correct_user_sum'].fillna(0,inplace=True)\n\n reg['answered_correctly_user'] = ((reg['answered_correctly_user']*reg['User_questions'])+reg['answered_correctly'])/(reg['User_questions']+1)\n reg['answer_wrong_user'] = 1-reg['answered_correctly_user']\n #reg['wrong_user_sum'] = reg['wrong_user_sum'] + reg['answered_wrong']\n #reg['correct_user_sum'] = reg['correct_user_sum'] + reg['answered_correctly']\n reg['User_questions'] = reg['User_questions']+1\n\n user_questions.append(reg[['user_id','User_questions']].to_pandas())\n correct_user.append(reg[['user_id','answered_correctly_user']].to_pandas())\n answer_wrong_user.append(reg[['user_id','answer_wrong_user']].to_pandas())\n #corrrect_user_sum.append(reg[[\"correct_user_sum\",'user_id']].to_pandas())\n #wrong_user_sum.append(reg[[\"wrong_user_sum\",'user_id']].to_pandas())\n\n user_questions.drop_duplicates(subset=['user_id'], keep='last')\n correct_user.drop_duplicates(subset=['user_id'], keep='last')\n answer_wrong_user.drop_duplicates(subset=['user_id'], keep='last')\n #corrrect_user_sum.drop_duplicates(subset=['user_id'], keep='last')\n #wrong_user_sum.drop_duplicates(subset=['user_id'], keep='last')\nHave a look at https://www.kaggle.com/its7171/lgbm-with-loop-feature-engineering for ideas on how to achieve something similar to this. I think count_u would be the same as your correct_user_sum
X['correct_user_sum']-1 this may not really prevent leakage
", "rawMarkdown": "X['correct_user_sum']-1 this may not really prevent leakage", "votes": null }, { "id": "1075595", "postDate": "11/11/2020 19:56:23", "content": "I suspect it as well. Tho I don't know why. Since its cumsum and done by row, it should theoretically remove answers from the same row. Even looking at the data frame it looks ok gives the same result if I use a shift(1)
", "rawMarkdown": "I suspect it as well. Tho I don't know why. Since its cumsum and done by row, it should theoretically remove answers from the same row. Even looking at the data frame it looks ok gives the same result if I use a shift(1)", "votes": null }, { "id": "1090501", "postDate": "11/25/2020 11:29:33", "content": "A leak free way to do this would be:
\nX[\"correct_user_sum\"] = X[[\"user_id\",'answered_correctly']].groupby([\"user_id\"])['answered_correctly'].apply(lambda x: x.cumsum().shift())\nDoing a X[\"correct_user_sum\"] = X[\"correct_user_sum\"] - 1 would still introduce a leak as correct_user_sum for that row would still be greater than previous correct_user_sum by 1 unit. I don't really know how a non-sequence model such as LGB could possibly exploit this leak.
Is this working within a reasonable time?
", "rawMarkdown": "Is this working within a reasonable time?", "votes": null }, { "id": "1150608", "postDate": "01/12/2021 17:45:19", "content": "Hi,
\nThanks. I have used some of the insights here:
\nhttps://www.kaggle.com/kritidoneria/beginner-wids21-feature-engineering-starter
I'm trying
\nNot sure about the feature importance yet.
", "votes": null, "replies": [ { "id": 1055027, "author_name": "pavelvpster", "author_url": "", "post_date": "10/20/2020 12:22:20", "content": "I am using Target Encoding for:
\nHow about seq2seq model
", "votes": null, "replies": [] }, { "id": 1055393, "author_name": "nyakaggle", "author_url": "", "post_date": "10/20/2020 18:12:38", "content": "\n\nFlag that shows that user watched at least one lecture
\n
You might also want to create one feature per lecture type:
\nI have 11 features:
\n['part_1_boolean',\n 'part_2_boolean',\n 'part_3_boolean',\n 'part_4_boolean',\n 'part_5_boolean',\n 'part_6_boolean',\n 'part_7_boolean',\n 'type_of_concept_boolean',\n 'type_of_intention_boolean',\n 'type_of_solving_question_boolean',\n 'type_of_starter_boolean']\nfor lecture types and for parts.
", "votes": null, "replies": [] } ] }, { "id": 1055598, "author_name": "suzukiseiya", "author_url": "", "post_date": "10/21/2020 00:12:59", "content": "I'm just trying to make features about congeniality between users and questions.
\nSpecifically, I compare the quesion's average correct answer rate for each \"part\" with the user's one.
\nI'm just trying now, so Not sure about the feature importance yet.
I am trying to split the users to categories: smart ones, with high percentage of correctly answered questions and not. It is also correlating with summary time spent on the platform by the user.
", "votes": null, "replies": [ { "id": 1056422, "author_name": "nyakaggle", "author_url": "", "post_date": "10/21/2020 17:26:52", "content": "@anaidashaginian Related to this, you might find this notebook interesting: https://www.kaggle.com/datafan07/riiid-challenge-eda-baseline-model#Answer-Accuracy---Time-Relations
", "votes": null, "replies": [] }, { "id": 1056427, "author_name": "kaushal2896", "author_url": "", "post_date": "10/21/2020 17:32:19", "content": "@anaidashaginian is that feature working for you? It's not working for me: https://www.kaggle.com/c/riiid-test-answer-prediction/discussion/190217
", "votes": null, "replies": [] } ] }, { "id": 1056355, "author_name": "anaidashaginian", "author_url": "", "post_date": "10/21/2020 16:14:37", "content": "Also I found the users which are both in train and test. I think, this could be useful.
", "votes": null, "replies": [ { "id": 1056748, "author_name": "msharuk589", "author_url": "", "post_date": "10/22/2020 03:35:50", "content": "@anaidashaginian Both in public and private test data??
", "votes": null, "replies": [] } ] }, { "id": 1060628, "author_name": "valentinart", "author_url": "", "post_date": "10/26/2020 12:18:02", "content": "What about trying to figure out a sort of \"learning curves\"? Something like a cumulative number of lectures or questions related to each part or some group of tags (they also can be clustered somehow).
\nI'm working on it now.
Will look into it.
", "votes": null, "replies": [] }, { "id": 1063734, "author_name": "alexj21", "author_url": "", "post_date": "10/29/2020 08:57:11", "content": "Does anyone manage to extract relevant information from the user's answer ?
", "votes": null, "replies": [ { "id": 1063749, "author_name": "pp2file", "author_url": "", "post_date": "10/29/2020 09:11:00", "content": "because testcase donot contain user's answer
", "votes": null, "replies": [] }, { "id": 1063754, "author_name": "rohanrao", "author_url": "", "post_date": "10/29/2020 09:19:36", "content": "Test data does contain user's answer in the prior_group_responses field.
I'm looking to update on the fly user's answer statistics in test set like some of us do for user answered correctly
", "votes": null, "replies": [] }, { "id": 1063765, "author_name": "alexj21", "author_url": "", "post_date": "10/29/2020 09:41:25", "content": "I'm curious about the team name \"question the answers\" in LB btw
", "votes": null, "replies": [] }, { "id": 1063978, "author_name": "adityaecdrid", "author_url": "", "post_date": "10/29/2020 14:58:05", "content": "\n\nI'm looking to update on the fly user's answer statistics in test set like some of us do for user answered correctly
\n
Had tried a simple decay on past feats but it didn't help as updating the old stats with new will have little effect on the old one unless you compute them over a good sample size is my current understanding .
", "votes": null, "replies": [] } ] }, { "id": 1065125, "author_name": "bturan19", "author_url": "", "post_date": "10/30/2020 21:48:24", "content": "I've implemented an expanding window average of target variable by user. I wrapped it with numba ant it runs all 100m rows in a minute.
\nimport numba\n\n@numba.jit(nopython=True)\ndef expanding_target_mean(data):\n \"\"\"\n it takes user_id and answered correctly columns\n and calculates expanding mean over target\n \"\"\"\n res = np.zeros_like(data[:,0])\n user_id = 0\n counter = 0\n for i in range(len(res)):\n if data[i, 0] != user_id:\n user_id = data[i, 0]\n res[i] = data[i, 1]\n counter = 0\n continue\n if data[i, 0] == user_id:\n counter += 1\n res[i] = np.mean(data[i-counter:i, 1])\n return res\n\ndata[\"expanding_target_mean\"] = expanding_target_mean(data[['user_id', 'answered_correctly']].astype(float).values)\nWhat is the res variable?
I have observed that content_id and bundle_id are the same thing if I am not missing something. So bundle_id does not add much?
", "votes": null, "replies": [ { "id": 1065462, "author_name": "rohanrao", "author_url": "", "post_date": "10/31/2020 10:10:34", "content": "They aren't always the same. Some examples:
\n
Thanks for the detailed table. It seems bundle id can take same value for different question ids but these different values are almost same as question id (<1%). Mathematically they are linearly correalated in the data, you can see it if you plot it. So not sure if it will add any additional info to the model.
", "votes": null, "replies": [] } ] }, { "id": 1069157, "author_name": "adityaecdrid", "author_url": "", "post_date": "11/04/2020 06:41:00", "content": "I feel t's crucial to engineer features which one can translate to the test set optimally as well..
", "votes": null, "replies": [] }, { "id": 1075040, "author_name": "iuryck", "author_url": "", "post_date": "11/11/2020 10:52:23", "content": "So I tried the following:
\nX[\"correct_user_sum\"] = X[[\"user_id\",'answered_correctly']].groupby([\"user_id\"])['answered_correctly'].cumsum()\nX['correct_user_sum'] = X['correct_user_sum']-1\nLast line is to remove leakage
\nThis feature performs really really well localy (around 0.80 AUC), but very very poorly on the LB, I have no idea why.
\nI've already tried different CV strategies and it always performs the same
How are you dealing with the new information and new users in the test set during submission?
", "votes": null, "replies": [] }, { "id": 1075062, "author_name": "iuryck", "author_url": "", "post_date": "11/11/2020 11:11:36", "content": "Tried simple merge and FillNa, then tried just a simple [correct_user_sum+answered_correctly]. Both cases gave the same result. All my features are put into Series, so it's easy to update and merge
", "votes": null, "replies": [] }, { "id": 1075065, "author_name": "iuryck", "author_url": "", "post_date": "11/11/2020 11:14:41", "content": "this is my update funtion that gets a Df with User-ID in previous loop and there answers in the next loop
\ndef registrar(df):\n reg = cudf.DataFrame()\n reg['user_id'] = df['user_id']\n reg['answered_correctly'] = df['answered_correctly']\n reg['answered_wrong'] = 1-reg['answered_correctly']\n #reg = cudf.merge(reg, cudf.DataFrame(wrong_user_sum[[\"wrong_user_sum\",'user_id']]), right_on='user_id',left_on='user_id',how='left')\n #reg = cudf.merge(reg, cudf.DataFrame(corrrect_user_sum[[\"correct_user_sum\",'user_id']]), right_on='user_id',left_on='user_id',how='left')\n reg = cudf.merge(reg, cudf.DataFrame(answer_wrong_user[['user_id','answer_wrong_user']]), right_on='user_id',left_on='user_id',how='left')\n reg = cudf.merge(reg, cudf.DataFrame(correct_user[['user_id','answered_correctly_user']]), right_on='user_id',left_on='user_id',how='left')\n reg = cudf.merge(reg, cudf.DataFrame(user_questions[['user_id','User_questions']]), right_on='user_id',left_on='user_id',how='left')\n\n reg['User_questions'].fillna(0,inplace=True)\n reg['answered_correctly_user'].fillna(np.float32(0.643215),inplace=True)\n reg['answer_wrong_user'].fillna(np.float32(1-0.643215),inplace=True)\n #reg['wrong_user_sum'].fillna(0,inplace=True)\n #reg['correct_user_sum'].fillna(0,inplace=True)\n\n reg['answered_correctly_user'] = ((reg['answered_correctly_user']*reg['User_questions'])+reg['answered_correctly'])/(reg['User_questions']+1)\n reg['answer_wrong_user'] = 1-reg['answered_correctly_user']\n #reg['wrong_user_sum'] = reg['wrong_user_sum'] + reg['answered_wrong']\n #reg['correct_user_sum'] = reg['correct_user_sum'] + reg['answered_correctly']\n reg['User_questions'] = reg['User_questions']+1\n\n user_questions.append(reg[['user_id','User_questions']].to_pandas())\n correct_user.append(reg[['user_id','answered_correctly_user']].to_pandas())\n answer_wrong_user.append(reg[['user_id','answer_wrong_user']].to_pandas())\n #corrrect_user_sum.append(reg[[\"correct_user_sum\",'user_id']].to_pandas())\n #wrong_user_sum.append(reg[[\"wrong_user_sum\",'user_id']].to_pandas())\n\n user_questions.drop_duplicates(subset=['user_id'], keep='last')\n correct_user.drop_duplicates(subset=['user_id'], keep='last')\n answer_wrong_user.drop_duplicates(subset=['user_id'], keep='last')\n #corrrect_user_sum.drop_duplicates(subset=['user_id'], keep='last')\n #wrong_user_sum.drop_duplicates(subset=['user_id'], keep='last')\nHave a look at https://www.kaggle.com/its7171/lgbm-with-loop-feature-engineering for ideas on how to achieve something similar to this. I think count_u would be the same as your correct_user_sum
X['correct_user_sum']-1 this may not really prevent leakage
", "votes": null, "replies": [] }, { "id": 1075595, "author_name": "iuryck", "author_url": "", "post_date": "11/11/2020 19:56:23", "content": "I suspect it as well. Tho I don't know why. Since its cumsum and done by row, it should theoretically remove answers from the same row. Even looking at the data frame it looks ok gives the same result if I use a shift(1)
", "votes": null, "replies": [] }, { "id": 1090501, "author_name": "doctorkael", "author_url": "", "post_date": "11/25/2020 11:29:33", "content": "A leak free way to do this would be:
\nX[\"correct_user_sum\"] = X[[\"user_id\",'answered_correctly']].groupby([\"user_id\"])['answered_correctly'].apply(lambda x: x.cumsum().shift())\nDoing a X[\"correct_user_sum\"] = X[\"correct_user_sum\"] - 1 would still introduce a leak as correct_user_sum for that row would still be greater than previous correct_user_sum by 1 unit. I don't really know how a non-sequence model such as LGB could possibly exploit this leak.
Is this working within a reasonable time?
", "votes": null, "replies": [] } ] }, { "id": 1150608, "author_name": "kritidoneria", "author_url": "", "post_date": "01/12/2021 17:45:19", "content": "Hi,
\nThanks. I have used some of the insights here:
\nhttps://www.kaggle.com/kritidoneria/beginner-wids21-feature-engineering-starter