{ "id": 582388, "title": "Chiming my thoughts on this competition", "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/582388", "author_name": "Ravi Ramakrishnan", "post_date": "2025-05-30T21:00:47.019000", "votes": 137, "comment_count": 32, "views": 0, "content": "

Hello all,

\n

I have participated in quite a few tabular challenges lately and have experience of success and failure in these types of assignments across my tryst with these challenges. I wish to chime a few thoughts regarding this competition in line with my participation in past editions of the challenge (PIU and Sleep States) and similar sized dataset challenges too, hoping these points will help a beginner orient himself/ herself well.

\n

Please feel free to agree/ disagree with these elements as they will surely help the reader orient him/ herself accordingly.

\n\n

Finally, have fun - this is a fun-learning filled challenge aimed at making one better at ML models and data wrangling. One learns the best when one enjoys the work done! Try and learn along and improve yourself - the public forums will evolve with time and this is a crucial factor in your final submission.

\n

All the best for a successful competition and happy learning!
\nBest regards,
\nRavi R

", "messages": [ { "id": 3214048, "postDate": "2025-05-30T21:00:47.020Z", "content": "

Hello all,

\n

I have participated in quite a few tabular challenges lately and have experience of success and failure in these types of assignments across my tryst with these challenges. I wish to chime a few thoughts regarding this competition in line with my participation in past editions of the challenge (PIU and Sleep States) and similar sized dataset challenges too, hoping these points will help a beginner orient himself/ herself well.

\n

Please feel free to agree/ disagree with these elements as they will surely help the reader orient him/ herself accordingly.

\n\n

Finally, have fun - this is a fun-learning filled challenge aimed at making one better at ML models and data wrangling. One learns the best when one enjoys the work done! Try and learn along and improve yourself - the public forums will evolve with time and this is a crucial factor in your final submission.

\n

All the best for a successful competition and happy learning!
\nBest regards,
\nRavi R

", "rawMarkdown": "Hello all,\n\nI have participated in quite a few tabular challenges lately and have experience of success and failure in these types of assignments across my tryst with these challenges. I wish to chime a few thoughts regarding this competition in line with my participation in past editions of the challenge (PIU and Sleep States) and similar sized dataset challenges too, hoping these points will help a beginner orient himself/ herself well.\n\nPlease feel free to agree/ disagree with these elements as they will surely help the reader orient him/ herself accordingly. \n\n- The dataset appears huge with a 1.2 GB size but we have approx. 5100 sequences of data here in the training set. This translates to an effective data size of 5100 rows, making this a **small data challenge**\n- Past CMI competitions, especially the PIU challenge was known for its shakeups. Single models and very simple approaches prevailed while large and complex ensembles failed - this could be a directive for this competition too\n- The metric is **prone to shakeup** as a few well classified points will alter the score considerably while a few poor classifications could reduce the score considerably. **Be careful to design simple models that generalize well rather than a lot of complex models with questionable generalization potential**\n- This is a code competition with an API - beginners transitioning with the playground competitions may find the API a bit different. Please follow the [standard submission kernel](https://www.kaggle.com/code/sohier/cmi-2025-demo-submission) as a guideline and edit this for starters\n- Remember to **test the prediction values on a dummy train set** before submitting, especially if you are new to this type of competitions - **erroneous submissions are counted in the daily quota**, so be careful and submit correctly \n- **Non-CV backed blends and blind fuses rarely work in such assignments** - stick to optimizing the CV scheme. You have 3 months going ahead and I am sure this competition does not require that long time to complete\n- Usually such competitions will attract a lot of participants and we could witness 4500+ teams at the end of the challenge as well. **A good way to approach such competitions is to set one's goals well and try and realize them as one moves along**. Usually setting easy goals and achieving them all along is better than setting a far sighted and difficult goal that could become difficult to realize.\n- Plan your time in August 2025 now and try and gauge if you could make time for this competition - **this competition will conclude on Sep2, so your time in August is crucial**\n- Start off with simple single models and make way for FE - this is crucial here. We don't need a baroque ensemble here- simple ML models seem to be sufficient if done well. Usually, leaders and toppers focus on FE in such competitions while the rest focus on model training - I am sure one's direction is clear here. \n- Finally, keep **track of your work with any experiment tracking tool.** It could be a simple excel sheet, a tool like W&B, Neptune.ai, MLOps tools like Kedro / MLFlow, etc. Experiment tracking is a good way to orient yourself with regard to ideas that worked and did not work and revisit your work later in the competition. Usually these artefacts are useful while teaming up as well as one could articulate the work done with aplomb.\n- **Save the fitted models and OOF predictions as you move along and keep them handy for later use - this trick will surely help you at the end of the competition when you may want to blend and fuse submissions and models for just an extra bit of alpha**\n\nFinally, have fun - this is a fun-learning filled challenge aimed at making one better at ML models and data wrangling. One learns the best when one enjoys the work done! Try and learn along and improve yourself - the public forums will evolve with time and this is a crucial factor in your final submission. \n\nAll the best for a successful competition and happy learning!\nBest regards,\nRavi R", "votes": 136 }, { "id": 3214073, "postDate": "2025-05-30T22:13:50.360Z", "content": "
\n
\n

The dataset appears huge with a 1.2 GB

\n
\n
\n

'huge'.
\nMeanwhile Geophysical Waveform data: ~500 GB 🤣
\nI wish it really was huge…small data is too shaky, I don't like that…

", "rawMarkdown": ">>The dataset appears huge with a 1.2 GB \n\n'huge'. \nMeanwhile Geophysical Waveform data: ~500 GB 🤣 \nI wish it really was huge...small data is too shaky, I don't like that... ", "votes": 5, "replies": [ { "id": 3214196, "postDate": "2025-05-31T05:04:49.420Z", "rawMarkdown": "", "votes": -1, "isDeleted": true, "replies": [ { "id": 3214203, "postDate": "2025-05-31T05:30:44.870Z", "content": "

Sure, sure, it just made me chuckle.
\nI wish I had sposored resources too 🤣

", "rawMarkdown": "Sure, sure, it just made me chuckle.\nI wish I had sposored resources too 🤣\n" } ] } ] }, { "id": 3215577, "postDate": "2025-06-02T12:45:58.197Z", "content": "

@ravi20076 one of the problem with this type of competitions is towards the end of competition someone shares a kernel with good public score and lot of participants just blindly fork and submit that code this can lead to shakeups as good chunk of competitors are overfitting rather than solving problem at the hand

", "rawMarkdown": "@ravi20076 one of the problem with this type of competitions is towards the end of competition someone shares a kernel with good public score and lot of participants just blindly fork and submit that code this can lead to shakeups as good chunk of competitors are overfitting rather than solving problem at the hand", "votes": 2 }, { "id": 3218840, "postDate": "2025-06-06T19:18:22.940Z", "content": "

This is a great set of insights. I completely agree that simplicity often leads to the best solutions, particularly in a competition like this where the data size can be misleading. It’s easy to overcomplicate things, but focusing on feature engineering and straightforward models can work wonders.

\n

The tip about testing on a dummy train set is great. It’s a small step that can save a lot of time and effort later on.

\n

Thanks for sharing these thoughtful tips.

", "rawMarkdown": "This is a great set of insights. I completely agree that simplicity often leads to the best solutions, particularly in a competition like this where the data size can be misleading. It’s easy to overcomplicate things, but focusing on feature engineering and straightforward models can work wonders.\n\nThe tip about testing on a dummy train set is great. It’s a small step that can save a lot of time and effort later on.\n\nThanks for sharing these thoughtful tips.", "votes": 1 }, { "id": 3246735, "postDate": "2025-07-11T12:34:03.610Z", "content": "

I had exactly the same thoughts with you for this competition. But unfortunately I face memory issues even with applying a simple logistic regression model. Any help about that? What are the main steps you follow to handle this?

", "rawMarkdown": "I had exactly the same thoughts with you for this competition. But unfortunately I face memory issues even with applying a simple logistic regression model. Any help about that? What are the main steps you follow to handle this? ", "replies": [ { "id": 3250341, "postDate": "2025-07-18T06:15:33.277Z", "rawMarkdown": "", "votes": -1, "isDeleted": true } ] }, { "id": 3240806, "postDate": "2025-07-04T08:34:18.900Z", "content": "

Thank you very much for your guidance. I will proceed accordingly and focus on building a strong foundation with simple models and well-thought-out feature engineering.

", "rawMarkdown": "Thank you very much for your guidance. I will proceed accordingly and focus on building a strong foundation with simple models and well-thought-out feature engineering.\n" }, { "id": 3217427, "postDate": "2025-06-05T02:17:15.573Z", "content": "

@ravi20076 , very useful advise , thanks for your experience sharing, After reading your article, I know how to start is better!

", "rawMarkdown": "@ravi20076 , very useful advise , thanks for your experience sharing, After reading your article, I know how to start is better!", "votes": 1 }, { "id": 3222517, "postDate": "2025-06-12T08:32:36.687Z", "content": "

Thank you very much for your recommendations. I am sure they will be of great help 😉

", "rawMarkdown": "Thank you very much for your recommendations. I am sure they will be of great help 😉" }, { "id": 3221971, "postDate": "2025-06-11T16:36:01.520Z", "content": "

Really appreciate your advice

", "rawMarkdown": "Really appreciate your advice " }, { "id": 3218468, "postDate": "2025-06-06T08:01:41.717Z", "content": "

Absolutely agree! These wild score swings can be so demoralizing — you put in tons of work and the score barely moves. But I guess that’s what makes these challenges a real test of patience and growth.

", "rawMarkdown": "Absolutely agree! These wild score swings can be so demoralizing — you put in tons of work and the score barely moves. But I guess that’s what makes these challenges a real test of patience and growth." }, { "id": 3216981, "postDate": "2025-06-04T11:26:55.660Z", "content": "

Hi!
\nThank you so much for your valuable insights on navigating this competition, especially your emphasis on 'simple models that generalize well' and the observation that this is effectively a 'small data challenge' with only ~5100 sequences.
\nThis is my first international competition. Considering the points you have mentioned, I'm curious about your perspective on using deep learning models for this competition. Do you think a DL approach would be suitable here, or would it fall into the category of 'complex models with questionable generalization potential' that you advise against?

", "rawMarkdown": "Hi!\nThank you so much for your valuable insights on navigating this competition, especially your emphasis on 'simple models that generalize well' and the observation that this is effectively a 'small data challenge' with only ~5100 sequences.\nThis is my first international competition. Considering the points you have mentioned, I'm curious about your perspective on using deep learning models for this competition. Do you think a DL approach would be suitable here, or would it fall into the category of 'complex models with questionable generalization potential' that you advise against? ", "replies": [ { "id": 3217003, "postDate": "2025-06-04T11:57:47.020Z", "content": "

Of course you can use deep learning - this post is to remind one of the pitfalls of blind blends and very complex approaches that don't often generalise well @abdulkareemomer

", "rawMarkdown": "Of course you can use deep learning - this post is to remind one of the pitfalls of blind blends and very complex approaches that don't often generalise well @abdulkareemomer ", "replies": [ { "id": 3217020, "postDate": "2025-06-04T12:27:10.447Z", "content": "

Thank you for the clarification.

", "rawMarkdown": "Thank you for the clarification." } ] }, { "id": 3218068, "postDate": "2025-06-05T18:40:05.923Z", "content": "

I'd go for simpler architectures though. From what @ravi20076 points out, shakeup can be brutal if your architecture is too fancy which will increase the odds of overfitting the LB

", "rawMarkdown": "I'd go for simpler architectures though. From what @ravi20076 points out, shakeup can be brutal if your architecture is too fancy which will increase the odds of overfitting the LB" } ] }, { "id": 3214097, "postDate": "2025-05-30T23:45:27.483Z", "content": "

I see some familiar faces on the discussions then I know this comp will have 4500+ participants 💀

", "rawMarkdown": "I see some familiar faces on the discussions then I know this comp will have 4500+ participants 💀", "votes": 2 }, { "id": 3216622, "postDate": "2025-06-03T20:57:56.187Z", "content": "

Thanks for the insights! First time in a non-playground competition so it was helpful to get a better understanding of this one! Good luck!

", "rawMarkdown": "Thanks for the insights! First time in a non-playground competition so it was helpful to get a better understanding of this one! Good luck!" }, { "id": 3215682, "postDate": "2025-06-02T14:42:46.320Z", "content": "

Thank you for sharing your perspective. :)
\nGiven your experience with similar datasets, would you recommend increasing the dataset size through feature engineering techniques? Would it help create a more substantial dataset for training the machine learning model on customized data?

", "rawMarkdown": "Thank you for sharing your perspective. :)\nGiven your experience with similar datasets, would you recommend increasing the dataset size through feature engineering techniques? Would it help create a more substantial dataset for training the machine learning model on customized data?" }, { "id": 3215197, "postDate": "2025-06-01T18:44:30.317Z", "content": "

Helpful insight….for beginners
\nthanks for sharing

", "rawMarkdown": "Helpful insight....for beginners\nthanks for sharing" }, { "id": 3214635, "postDate": "2025-05-31T23:01:48.240Z", "content": "

I've also noticed that in most competitions, simple models tend to prevail, which is fascinating.

", "rawMarkdown": "I've also noticed that in most competitions, simple models tend to prevail, which is fascinating." }, { "id": 3220953, "postDate": "2025-06-10T07:33:42.347Z", "content": "

Thank you for your insights and recommendations.

", "rawMarkdown": "Thank you for your insights and recommendations.", "votes": -1 }, { "id": 3220456, "postDate": "2025-06-09T10:47:56.107Z", "content": "

That was a great and comprehensive analysis, thank you very much.
\nA small question, I currently see 8151 unique sequence_ids in the training dataset, while you mentioned about 5100 sequences. Could this difference be due to different versions of the dataset? Or maybe I'm doing something wrong!?

", "rawMarkdown": "That was a great and comprehensive analysis, thank you very much.\nA small question, I currently see 8151 unique sequence_ids in the training dataset, while you mentioned about 5100 sequences. Could this difference be due to different versions of the dataset? Or maybe I'm doing something wrong!?\n", "votes": -1 }, { "id": 3220158, "postDate": "2025-06-08T22:52:24.370Z", "content": "

Some really great advice here, thanks Ravi.

", "rawMarkdown": "Some really great advice here, thanks Ravi.", "votes": -1 }, { "id": 3215107, "postDate": "2025-06-01T15:39:41.410Z", "content": "

It’s a very helpful insight for beginners.. Thanks!

", "rawMarkdown": "It’s a very helpful insight for beginners.. Thanks!", "votes": 1, "isDeleted": true }, { "id": 3243310, "postDate": "2025-07-07T02:26:50.537Z", "content": "

Thanks for your guidance.

", "rawMarkdown": "Thanks for your guidance.", "votes": 1 }, { "id": 3220839, "postDate": "2025-06-10T03:35:04.213Z", "content": "

Thanks! It's very useful!

", "rawMarkdown": "Thanks! It's very useful!", "votes": 1 }, { "id": 3220617, "postDate": "2025-06-09T15:48:10.473Z", "content": "

Thanks a lot for your sharing.

", "rawMarkdown": "Thanks a lot for your sharing.", "votes": 1 }, { "id": 3218969, "postDate": "2025-06-07T01:54:36.487Z", "content": "

good stuff!

", "rawMarkdown": "good stuff!", "votes": 1 }, { "id": 3218491, "postDate": "2025-06-06T08:52:16.107Z", "content": "

thanks for sharing your knowledge

", "rawMarkdown": "thanks for sharing your knowledge ", "votes": 1 }, { "id": 3214513, "postDate": "2025-05-31T17:25:37.153Z", "content": "

Thanks, pretty useful tips.

", "rawMarkdown": "Thanks, pretty useful tips.", "votes": 1 }, { "id": 3214217, "postDate": "2025-05-31T05:57:31.073Z", "content": "

Thanks! It's cool

", "rawMarkdown": "Thanks! It's cool\n", "votes": 1 } ], "comments": [ { "id": 3214073, "author_name": "greySnow", "author_url": "", "post_date": "2025-05-30T22:13:50.360000", "content": "
\n
\n

The dataset appears huge with a 1.2 GB

\n
\n
\n

'huge'.
\nMeanwhile Geophysical Waveform data: ~500 GB 🤣
\nI wish it really was huge…small data is too shaky, I don't like that…

", "votes": 5, "replies": [ { "id": 3214196, "author_name": "", "author_url": "", "post_date": "2025-05-31T05:04:49.420000", "content": "", "votes": -1, "replies": [ { "id": 3214203, "author_name": "greySnow", "author_url": "", "post_date": "2025-05-31T05:30:44.870000", "content": "

Sure, sure, it just made me chuckle.
\nI wish I had sposored resources too 🤣

", "votes": 0, "replies": [] } ] } ] }, { "id": 3215577, "author_name": "Athar Sayed", "author_url": "", "post_date": "2025-06-02T12:45:58.197000", "content": "

@ravi20076 one of the problem with this type of competitions is towards the end of competition someone shares a kernel with good public score and lot of participants just blindly fork and submit that code this can lead to shakeups as good chunk of competitors are overfitting rather than solving problem at the hand

", "votes": 2, "replies": [] }, { "id": 3218840, "author_name": "Sourav Basak Shuvo", "author_url": "", "post_date": "2025-06-06T19:18:22.940000", "content": "

This is a great set of insights. I completely agree that simplicity often leads to the best solutions, particularly in a competition like this where the data size can be misleading. It’s easy to overcomplicate things, but focusing on feature engineering and straightforward models can work wonders.

\n

The tip about testing on a dummy train set is great. It’s a small step that can save a lot of time and effort later on.

\n

Thanks for sharing these thoughtful tips.

", "votes": 1, "replies": [] }, { "id": 3246735, "author_name": "Aristotelis Patsopoulos", "author_url": "", "post_date": "2025-07-11T12:34:03.610000", "content": "

I had exactly the same thoughts with you for this competition. But unfortunately I face memory issues even with applying a simple logistic regression model. Any help about that? What are the main steps you follow to handle this?

", "votes": 0, "replies": [ { "id": 3250341, "author_name": "", "author_url": "", "post_date": "2025-07-18T06:15:33.277000", "content": "", "votes": -1, "replies": [] } ] }, { "id": 3240806, "author_name": "Utkarsh Mavi", "author_url": "", "post_date": "2025-07-04T08:34:18.900000", "content": "

Thank you very much for your guidance. I will proceed accordingly and focus on building a strong foundation with simple models and well-thought-out feature engineering.

", "votes": 0, "replies": [] }, { "id": 3217427, "author_name": "Chris_Shen", "author_url": "", "post_date": "2025-06-05T02:17:15.573000", "content": "

@ravi20076 , very useful advise , thanks for your experience sharing, After reading your article, I know how to start is better!

", "votes": 1, "replies": [] }, { "id": 3222517, "author_name": "Ricardo Pirabed A.A.", "author_url": "", "post_date": "2025-06-12T08:32:36.687000", "content": "

Thank you very much for your recommendations. I am sure they will be of great help 😉

", "votes": 0, "replies": [] }, { "id": 3221971, "author_name": "yousef Elshahat", "author_url": "", "post_date": "2025-06-11T16:36:01.520000", "content": "

Really appreciate your advice

", "votes": 0, "replies": [] }, { "id": 3218468, "author_name": "Sisyphus", "author_url": "", "post_date": "2025-06-06T08:01:41.717000", "content": "

Absolutely agree! These wild score swings can be so demoralizing — you put in tons of work and the score barely moves. But I guess that’s what makes these challenges a real test of patience and growth.

", "votes": 0, "replies": [] }, { "id": 3216981, "author_name": "Abdulkareem Omer", "author_url": "", "post_date": "2025-06-04T11:26:55.660000", "content": "

Hi!
\nThank you so much for your valuable insights on navigating this competition, especially your emphasis on 'simple models that generalize well' and the observation that this is effectively a 'small data challenge' with only ~5100 sequences.
\nThis is my first international competition. Considering the points you have mentioned, I'm curious about your perspective on using deep learning models for this competition. Do you think a DL approach would be suitable here, or would it fall into the category of 'complex models with questionable generalization potential' that you advise against?

", "votes": 0, "replies": [ { "id": 3217003, "author_name": "Ravi Ramakrishnan", "author_url": "", "post_date": "2025-06-04T11:57:47.020000", "content": "

Of course you can use deep learning - this post is to remind one of the pitfalls of blind blends and very complex approaches that don't often generalise well @abdulkareemomer

", "votes": 0, "replies": [ { "id": 3217020, "author_name": "Abdulkareem Omer", "author_url": "", "post_date": "2025-06-04T12:27:10.447000", "content": "

Thank you for the clarification.

", "votes": 0, "replies": [] } ] }, { "id": 3218068, "author_name": "Victor S", "author_url": "", "post_date": "2025-06-05T18:40:05.923000", "content": "

I'd go for simpler architectures though. From what @ravi20076 points out, shakeup can be brutal if your architecture is too fancy which will increase the odds of overfitting the LB

", "votes": 0, "replies": [] } ] }, { "id": 3214097, "author_name": "Tom", "author_url": "", "post_date": "2025-05-30T23:45:27.483000", "content": "

I see some familiar faces on the discussions then I know this comp will have 4500+ participants 💀

", "votes": 2, "replies": [] }, { "id": 3216622, "author_name": "Márk Antal Csizmadia", "author_url": "", "post_date": "2025-06-03T20:57:56.187000", "content": "

Thanks for the insights! First time in a non-playground competition so it was helpful to get a better understanding of this one! Good luck!

", "votes": 0, "replies": [] }, { "id": 3215682, "author_name": "Ishita", "author_url": "", "post_date": "2025-06-02T14:42:46.320000", "content": "

Thank you for sharing your perspective. :)
\nGiven your experience with similar datasets, would you recommend increasing the dataset size through feature engineering techniques? Would it help create a more substantial dataset for training the machine learning model on customized data?

", "votes": 0, "replies": [] }, { "id": 3215197, "author_name": "Sarah Arshad", "author_url": "", "post_date": "2025-06-01T18:44:30.317000", "content": "

Helpful insight….for beginners
\nthanks for sharing

", "votes": 0, "replies": [] }, { "id": 3214635, "author_name": "Michael R", "author_url": "", "post_date": "2025-05-31T23:01:48.240000", "content": "

I've also noticed that in most competitions, simple models tend to prevail, which is fascinating.

", "votes": 0, "replies": [] }, { "id": 3220953, "author_name": "Bilal A. Chaudhry", "author_url": "", "post_date": "2025-06-10T07:33:42.347000", "content": "

Thank you for your insights and recommendations.

", "votes": -1, "replies": [] }, { "id": 3220456, "author_name": "Mando", "author_url": "", "post_date": "2025-06-09T10:47:56.107000", "content": "

That was a great and comprehensive analysis, thank you very much.
\nA small question, I currently see 8151 unique sequence_ids in the training dataset, while you mentioned about 5100 sequences. Could this difference be due to different versions of the dataset? Or maybe I'm doing something wrong!?

", "votes": -1, "replies": [] }, { "id": 3220158, "author_name": "Nick Potter", "author_url": "", "post_date": "2025-06-08T22:52:24.370000", "content": "

Some really great advice here, thanks Ravi.

", "votes": -1, "replies": [] }, { "id": 3215107, "author_name": "", "author_url": "", "post_date": "2025-06-01T15:39:41.410000", "content": "

It’s a very helpful insight for beginners.. Thanks!

", "votes": 1, "replies": [] }, { "id": 3243310, "author_name": "backpack", "author_url": "", "post_date": "2025-07-07T02:26:50.537000", "content": "

Thanks for your guidance.

", "votes": 1, "replies": [] }, { "id": 3220839, "author_name": "Kazukichi", "author_url": "", "post_date": "2025-06-10T03:35:04.213000", "content": "

Thanks! It's very useful!

", "votes": 1, "replies": [] }, { "id": 3220617, "author_name": "DriveBeyond", "author_url": "", "post_date": "2025-06-09T15:48:10.473000", "content": "

Thanks a lot for your sharing.

", "votes": 1, "replies": [] }, { "id": 3218969, "author_name": "roshan", "author_url": "", "post_date": "2025-06-07T01:54:36.487000", "content": "

good stuff!

", "votes": 1, "replies": [] }, { "id": 3218491, "author_name": "raman chaudhary jarodiya", "author_url": "", "post_date": "2025-06-06T08:52:16.107000", "content": "

thanks for sharing your knowledge

", "votes": 1, "replies": [] }, { "id": 3214513, "author_name": "Sandeep M", "author_url": "", "post_date": "2025-05-31T17:25:37.153000", "content": "

Thanks, pretty useful tips.

", "votes": 1, "replies": [] }, { "id": 3214217, "author_name": "qdssad", "author_url": "", "post_date": "2025-05-31T05:57:31.073000", "content": "

Thanks! It's cool

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "3214048": "Hello all,\n\nI have participated in quite a few tabular challenges lately and have experience of success and failure in these types of assignments across my tryst with these challenges. I wish to chime a few thoughts regarding this competition in line with my participation in past editions of the challenge (PIU and Sleep States) and similar sized dataset challenges too, hoping these points will help a beginner orient himself/ herself well.\n\nPlease feel free to agree/ disagree with these elements as they will surely help the reader orient him/ herself accordingly. \n\n- The dataset appears huge with a 1.2 GB size but we have approx. 5100 sequences of data here in the training set. This translates to an effective data size of 5100 rows, making this a **small data challenge**\n- Past CMI competitions, especially the PIU challenge was known for its shakeups. Single models and very simple approaches prevailed while large and complex ensembles failed - this could be a directive for this competition too\n- The metric is **prone to shakeup** as a few well classified points will alter the score considerably while a few poor classifications could reduce the score considerably. **Be careful to design simple models that generalize well rather than a lot of complex models with questionable generalization potential**\n- This is a code competition with an API - beginners transitioning with the playground competitions may find the API a bit different. Please follow the [standard submission kernel](https://www.kaggle.com/code/sohier/cmi-2025-demo-submission) as a guideline and edit this for starters\n- Remember to **test the prediction values on a dummy train set** before submitting, especially if you are new to this type of competitions - **erroneous submissions are counted in the daily quota**, so be careful and submit correctly \n- **Non-CV backed blends and blind fuses rarely work in such assignments** - stick to optimizing the CV scheme. You have 3 months going ahead and I am sure this competition does not require that long time to complete\n- Usually such competitions will attract a lot of participants and we could witness 4500+ teams at the end of the challenge as well. **A good way to approach such competitions is to set one's goals well and try and realize them as one moves along**. Usually setting easy goals and achieving them all along is better than setting a far sighted and difficult goal that could become difficult to realize.\n- Plan your time in August 2025 now and try and gauge if you could make time for this competition - **this competition will conclude on Sep2, so your time in August is crucial**\n- Start off with simple single models and make way for FE - this is crucial here. We don't need a baroque ensemble here- simple ML models seem to be sufficient if done well. Usually, leaders and toppers focus on FE in such competitions while the rest focus on model training - I am sure one's direction is clear here. \n- Finally, keep **track of your work with any experiment tracking tool.** It could be a simple excel sheet, a tool like W&B, Neptune.ai, MLOps tools like Kedro / MLFlow, etc. Experiment tracking is a good way to orient yourself with regard to ideas that worked and did not work and revisit your work later in the competition. Usually these artefacts are useful while teaming up as well as one could articulate the work done with aplomb.\n- **Save the fitted models and OOF predictions as you move along and keep them handy for later use - this trick will surely help you at the end of the competition when you may want to blend and fuse submissions and models for just an extra bit of alpha**\n\nFinally, have fun - this is a fun-learning filled challenge aimed at making one better at ML models and data wrangling. One learns the best when one enjoys the work done! Try and learn along and improve yourself - the public forums will evolve with time and this is a crucial factor in your final submission. \n\nAll the best for a successful competition and happy learning!\nBest regards,\nRavi R", "3214073": ">>The dataset appears huge with a 1.2 GB \n\n'huge'. \nMeanwhile Geophysical Waveform data: ~500 GB 🤣 \nI wish it really was huge...small data is too shaky, I don't like that... ", "3215577": "@ravi20076 one of the problem with this type of competitions is towards the end of competition someone shares a kernel with good public score and lot of participants just blindly fork and submit that code this can lead to shakeups as good chunk of competitors are overfitting rather than solving problem at the hand", "3218840": "This is a great set of insights. I completely agree that simplicity often leads to the best solutions, particularly in a competition like this where the data size can be misleading. It’s easy to overcomplicate things, but focusing on feature engineering and straightforward models can work wonders.\n\nThe tip about testing on a dummy train set is great. It’s a small step that can save a lot of time and effort later on.\n\nThanks for sharing these thoughtful tips.", "3246735": "I had exactly the same thoughts with you for this competition. But unfortunately I face memory issues even with applying a simple logistic regression model. Any help about that? What are the main steps you follow to handle this? ", "3240806": "Thank you very much for your guidance. I will proceed accordingly and focus on building a strong foundation with simple models and well-thought-out feature engineering.\n", "3217427": "@ravi20076 , very useful advise , thanks for your experience sharing, After reading your article, I know how to start is better!", "3222517": "Thank you very much for your recommendations. I am sure they will be of great help 😉", "3221971": "Really appreciate your advice ", "3218468": "Absolutely agree! These wild score swings can be so demoralizing — you put in tons of work and the score barely moves. But I guess that’s what makes these challenges a real test of patience and growth.", "3216981": "Hi!\nThank you so much for your valuable insights on navigating this competition, especially your emphasis on 'simple models that generalize well' and the observation that this is effectively a 'small data challenge' with only ~5100 sequences.\nThis is my first international competition. Considering the points you have mentioned, I'm curious about your perspective on using deep learning models for this competition. Do you think a DL approach would be suitable here, or would it fall into the category of 'complex models with questionable generalization potential' that you advise against? ", "3214097": "I see some familiar faces on the discussions then I know this comp will have 4500+ participants 💀", "3216622": "Thanks for the insights! First time in a non-playground competition so it was helpful to get a better understanding of this one! Good luck!", "3215682": "Thank you for sharing your perspective. :)\nGiven your experience with similar datasets, would you recommend increasing the dataset size through feature engineering techniques? Would it help create a more substantial dataset for training the machine learning model on customized data?", "3215197": "Helpful insight....for beginners\nthanks for sharing", "3214635": "I've also noticed that in most competitions, simple models tend to prevail, which is fascinating.", "3220953": "Thank you for your insights and recommendations.", "3220456": "That was a great and comprehensive analysis, thank you very much.\nA small question, I currently see 8151 unique sequence_ids in the training dataset, while you mentioned about 5100 sequences. Could this difference be due to different versions of the dataset? Or maybe I'm doing something wrong!?\n", "3220158": "Some really great advice here, thanks Ravi.", "3215107": "It’s a very helpful insight for beginners.. Thanks!", "3243310": "Thanks for your guidance.", "3220839": "Thanks! It's very useful!", "3220617": "Thanks a lot for your sharing.", "3218969": "good stuff!", "3218491": "thanks for sharing your knowledge ", "3214513": "Thanks, pretty useful tips.", "3214217": "Thanks! It's cool\n" } }