Hi all, it seems our time is running out! In the Kaggle spirit of learning I thought I would share some of what I learned with you all. Please do remember to not focus on the final LB score, but what you gained in knowledge from this competition! I would also like to hear some of what you all learned in this comp?
\nUsing 1d/2d CNNs for time series - This strategy later changed the game with popular public notebooks but it is very good for segmentation of seemingly signaled data. These practices are also used in other industries like finance!
Getting more familiar with python scripting - It sounds goofy to say it out loud but I really am not all that familiar with script, virtually all I have worked with has been notebooks so being able to dive into @tubotubo ’s popular work in order to search for growth opportunity was a rude awakening and a learning experience for me
Getting familiar with Git and commands - another one that seems odd but git is one of those things that your really only understand it if you use it frequently and I have certainly not been one to use it often in the past, but now I may start!
Multi dimensional time series analysis - in this project we had to be able to process time series data with great precision, over the years I have grown less fond of time series data because of how famously difficult it can be for what seems so simple to us. Being able to dynamically analyze sleep data by series_id made it much easier for me make features in later models that would give me potential for higher scores.
Signal processing- Signal processing was really big for this competition because we had to be able to try to filter out false positives as well as try to identify where the data tracker was off the child. I come from a physics background so this was familiar territory for me, that being said it has been awhile and it was nice to revisit it in a meaningful way.
Time series ensembling - the first time I have ensembled a time series model, and though my final iterations did not produce as much value as my partners aggregations I was forced to think through many ways to do this, simple average, weighted average by CV score, second stage ensembling (that did not work at all for this one), and more!
All in all this competition definitely gave me some struggles but I should be that much better for it! I hope many of you share the same beliefs, and good luck in the private LB!
", "messages": [ { "id": 2550214, "postDate": "2023-12-05T22:53:36.603Z", "content": "Hi all, it seems our time is running out! In the Kaggle spirit of learning I thought I would share some of what I learned with you all. Please do remember to not focus on the final LB score, but what you gained in knowledge from this competition! I would also like to hear some of what you all learned in this comp?
\nUsing 1d/2d CNNs for time series - This strategy later changed the game with popular public notebooks but it is very good for segmentation of seemingly signaled data. These practices are also used in other industries like finance!
Getting more familiar with python scripting - It sounds goofy to say it out loud but I really am not all that familiar with script, virtually all I have worked with has been notebooks so being able to dive into @tubotubo ’s popular work in order to search for growth opportunity was a rude awakening and a learning experience for me
Getting familiar with Git and commands - another one that seems odd but git is one of those things that your really only understand it if you use it frequently and I have certainly not been one to use it often in the past, but now I may start!
Multi dimensional time series analysis - in this project we had to be able to process time series data with great precision, over the years I have grown less fond of time series data because of how famously difficult it can be for what seems so simple to us. Being able to dynamically analyze sleep data by series_id made it much easier for me make features in later models that would give me potential for higher scores.
Signal processing- Signal processing was really big for this competition because we had to be able to try to filter out false positives as well as try to identify where the data tracker was off the child. I come from a physics background so this was familiar territory for me, that being said it has been awhile and it was nice to revisit it in a meaningful way.
Time series ensembling - the first time I have ensembled a time series model, and though my final iterations did not produce as much value as my partners aggregations I was forced to think through many ways to do this, simple average, weighted average by CV score, second stage ensembling (that did not work at all for this one), and more!
All in all this competition definitely gave me some struggles but I should be that much better for it! I hope many of you share the same beliefs, and good luck in the private LB!
", "rawMarkdown": "Hi all, it seems our time is running out! In the Kaggle spirit of learning I thought I would share some of what I learned with you all. Please do remember to not focus on the final LB score, but what you gained in knowledge from this competition! I would also like to hear some of what you all learned in this comp?\n\n1. Using 1d/2d CNNs for time series - This strategy later changed the game with popular public notebooks but it is very good for segmentation of seemingly signaled data. These practices are also used in other industries like finance!\n\n2. Getting more familiar with python scripting - It sounds goofy to say it out loud but I really am not all that familiar with script, virtually all I have worked with has been notebooks so being able to dive into @tubotubo ’s popular work in order to search for growth opportunity was a rude awakening and a learning experience for me\n\n3. Getting familiar with Git and commands - another one that seems odd but git is one of those things that your really only understand it if you use it frequently and I have certainly not been one to use it often in the past, but now I may start!\n\n4. Multi dimensional time series analysis - in this project we had to be able to process time series data with great precision, over the years I have grown less fond of time series data because of how famously difficult it can be for what seems so simple to us. Being able to dynamically analyze sleep data by series_id made it much easier for me make features in later models that would give me potential for higher scores.\n\n5. Signal processing- Signal processing was really big for this competition because we had to be able to try to filter out false positives as well as try to identify where the data tracker was off the child. I come from a physics background so this was familiar territory for me, that being said it has been awhile and it was nice to revisit it in a meaningful way.\n\n6. Time series ensembling - the first time I have ensembled a time series model, and though my final iterations did not produce as much value as my partners aggregations I was forced to think through many ways to do this, simple average, weighted average by CV score, second stage ensembling (that did not work at all for this one), and more!\n\nAll in all this competition definitely gave me some struggles but I should be that much better for it! I hope many of you share the same beliefs, and good luck in the private LB! ", "votes": 9 }, { "id": 2550367, "postDate": "2023-12-06T02:09:42.057Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2550367, "author_name": "", "author_url": "", "post_date": "2023-12-06T02:09:42.057000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2550214": "Hi all, it seems our time is running out! In the Kaggle spirit of learning I thought I would share some of what I learned with you all. Please do remember to not focus on the final LB score, but what you gained in knowledge from this competition! I would also like to hear some of what you all learned in this comp?\n\n1. Using 1d/2d CNNs for time series - This strategy later changed the game with popular public notebooks but it is very good for segmentation of seemingly signaled data. These practices are also used in other industries like finance!\n\n2. Getting more familiar with python scripting - It sounds goofy to say it out loud but I really am not all that familiar with script, virtually all I have worked with has been notebooks so being able to dive into @tubotubo ’s popular work in order to search for growth opportunity was a rude awakening and a learning experience for me\n\n3. Getting familiar with Git and commands - another one that seems odd but git is one of those things that your really only understand it if you use it frequently and I have certainly not been one to use it often in the past, but now I may start!\n\n4. Multi dimensional time series analysis - in this project we had to be able to process time series data with great precision, over the years I have grown less fond of time series data because of how famously difficult it can be for what seems so simple to us. Being able to dynamically analyze sleep data by series_id made it much easier for me make features in later models that would give me potential for higher scores.\n\n5. Signal processing- Signal processing was really big for this competition because we had to be able to try to filter out false positives as well as try to identify where the data tracker was off the child. I come from a physics background so this was familiar territory for me, that being said it has been awhile and it was nice to revisit it in a meaningful way.\n\n6. Time series ensembling - the first time I have ensembled a time series model, and though my final iterations did not produce as much value as my partners aggregations I was forced to think through many ways to do this, simple average, weighted average by CV score, second stage ensembling (that did not work at all for this one), and more!\n\nAll in all this competition definitely gave me some struggles but I should be that much better for it! I hope many of you share the same beliefs, and good luck in the private LB! ", "2550367": "" } }