Since actigraphy data is missing for the majority of the dataset, I am considering building two models: one that leverages actigraphy data and another without it. I wanted to get your thoughts on this approach and see if others have tried something similar. The idea is to then use separate models and inference them separately on test set for data points with and without actigraphy data.
", "messages": [ { "id": 3031731, "postDate": "2024-10-30T03:27:20.630Z", "content": "Since actigraphy data is missing for the majority of the dataset, I am considering building two models: one that leverages actigraphy data and another without it. I wanted to get your thoughts on this approach and see if others have tried something similar. The idea is to then use separate models and inference them separately on test set for data points with and without actigraphy data.
", "rawMarkdown": "Since actigraphy data is missing for the majority of the dataset, I am considering building two models: one that leverages actigraphy data and another without it. I wanted to get your thoughts on this approach and see if others have tried something similar. The idea is to then use separate models and inference them separately on test set for data points with and without actigraphy data. ", "votes": 4 }, { "id": 3035895, "postDate": "2024-11-04T01:27:33.200Z", "content": "Good point, I'll have to give it a try. Also, the 996 train ids with parquet files do not all have much clean data -- I've been focussing on about 780 of them that have at least a week of good data. In any case, fitting with-parquet ids by themselves could be a better way to clearly see if a custom parquet feature is useful.
", "rawMarkdown": "Good point, I'll have to give it a try. Also, the 996 train ids with parquet files do not all have much clean data -- I've been focussing on about 780 of them that have at least a week of good data. In any case, fitting with-parquet ids by themselves could be a better way to clearly see if a custom parquet feature is useful. ", "votes": 1 }, { "id": 3034151, "postDate": "2024-11-01T20:27:01.607Z", "content": "I think this is a valid approach and is something you should take into account. People far too quickly remove missing data while it is perfectly sensible to make 2 models, 1 model that creates a prediction when only the tabular data is available and another model that also incorporates the time-series data.
", "rawMarkdown": "I think this is a valid approach and is something you should take into account. People far too quickly remove missing data while it is perfectly sensible to make 2 models, 1 model that creates a prediction when only the tabular data is available and another model that also incorporates the time-series data. ", "votes": 1, "replies": [ { "id": 3034354, "postDate": "2024-11-02T03:58:25.003Z", "content": "yup, the amount of missing values especially with the actigraphy data encouraged me to go with this approach hope it works out
", "rawMarkdown": "yup, the amount of missing values especially with the actigraphy data encouraged me to go with this approach hope it works out" } ] }, { "id": 3034628, "postDate": "2024-11-02T12:13:46.820Z", "content": "I have also tried running models with and without actigraphy data. I have noted when actigraphy data is added it slightly improve the score on LeaderBoard. I will do some more experiments with it also in future.
", "rawMarkdown": "I have also tried running models with and without actigraphy data. I have noted when actigraphy data is added it slightly improve the score on LeaderBoard. I will do some more experiments with it also in future." }, { "id": 3032833, "postDate": "2024-10-31T12:57:42.483Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3035895, "author_name": "Daniel Dewey", "author_url": "", "post_date": "2024-11-04T01:27:33.200000", "content": "Good point, I'll have to give it a try. Also, the 996 train ids with parquet files do not all have much clean data -- I've been focussing on about 780 of them that have at least a week of good data. In any case, fitting with-parquet ids by themselves could be a better way to clearly see if a custom parquet feature is useful.
", "votes": 1, "replies": [] }, { "id": 3034151, "author_name": "Wouter Mostard", "author_url": "", "post_date": "2024-11-01T20:27:01.607000", "content": "I think this is a valid approach and is something you should take into account. People far too quickly remove missing data while it is perfectly sensible to make 2 models, 1 model that creates a prediction when only the tabular data is available and another model that also incorporates the time-series data.
", "votes": 1, "replies": [ { "id": 3034354, "author_name": "Deepak Saldanha", "author_url": "", "post_date": "2024-11-02T03:58:25.003000", "content": "yup, the amount of missing values especially with the actigraphy data encouraged me to go with this approach hope it works out
", "votes": 0, "replies": [] } ] }, { "id": 3034628, "author_name": "Taimour Nazar", "author_url": "", "post_date": "2024-11-02T12:13:46.820000", "content": "I have also tried running models with and without actigraphy data. I have noted when actigraphy data is added it slightly improve the score on LeaderBoard. I will do some more experiments with it also in future.
", "votes": 0, "replies": [] }, { "id": 3032833, "author_name": "", "author_url": "", "post_date": "2024-10-31T12:57:42.483000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3031731": "Since actigraphy data is missing for the majority of the dataset, I am considering building two models: one that leverages actigraphy data and another without it. I wanted to get your thoughts on this approach and see if others have tried something similar. The idea is to then use separate models and inference them separately on test set for data points with and without actigraphy data. ", "3035895": "Good point, I'll have to give it a try. Also, the 996 train ids with parquet files do not all have much clean data -- I've been focussing on about 780 of them that have at least a week of good data. In any case, fitting with-parquet ids by themselves could be a better way to clearly see if a custom parquet feature is useful. ", "3034151": "I think this is a valid approach and is something you should take into account. People far too quickly remove missing data while it is perfectly sensible to make 2 models, 1 model that creates a prediction when only the tabular data is available and another model that also incorporates the time-series data. ", "3034628": "I have also tried running models with and without actigraphy data. I have noted when actigraphy data is added it slightly improve the score on LeaderBoard. I will do some more experiments with it also in future.", "3032833": "" } }