I don't have too much experience with time series classification tasks so I decided to go with a pretty simple approach. I used tsfresh library to extract a set of features from time series data. Also, the library helps to pick only relevant features among extracted using null-hypothesis testing framework. Then I trained a LightGBM classifier using the extracted features.
This kernel shows an example of this approach. The kernel gives 0.59 on private. However, if you extract more features, pick only the relevant ones, tune the classifier parameters and use k-fold validation, you'll probably get something around 0.66. (I am attaching a submission file that gives me that score on private).
Though this solution is far away from the top positions, it is very straightforward and could be implemented within a couple of hours. So it seems to be a good baseline solution for time series classification problems.
", "comments": [ { "id": 515813, "author_name": "Prashanth Thangavel", "author_url": "", "post_date": "2019-04-13T07:00:01.287000", "content": "Congrats! Thanks for sharing the approach :) well deserved
", "votes": 1, "replies": [] }, { "id": 515241, "author_name": "danisamba", "author_url": "", "post_date": "2019-04-12T10:58:34.067000", "content": "Congrats for being in the top 100. I also got a shake up in the leaderboard with a similar approach.
", "votes": 1, "replies": [ { "id": 515260, "author_name": "Ilia Zaitsev", "author_url": "", "post_date": "2019-04-12T11:14:24.407000", "content": "Thank you! Yes, it seems that in the competitions with such a massive shake up the simple approaches show good results. I remember that I had some jump up in the recent Microsoft Malware competition as well, and the winners used relatively simple models too.
", "votes": 0, "replies": [] } ] } ], "comment_count": 3, "messages": [ { "id": 515813, "postDate": "2019-04-13T07:00:01.287Z", "content": "Congrats! Thanks for sharing the approach :) well deserved
", "rawMarkdown": "Congrats! Thanks for sharing the approach :) well deserved", "votes": 1 }, { "id": 515241, "postDate": "2019-04-12T10:58:34.067Z", "content": "Congrats for being in the top 100. I also got a shake up in the leaderboard with a similar approach.
", "rawMarkdown": "Congrats for being in the top 100. I also got a shake up in the leaderboard with a similar approach.", "votes": 1, "replies": [ { "id": 515260, "postDate": "2019-04-12T11:14:24.407Z", "content": "Thank you! Yes, it seems that in the competitions with such a massive shake up the simple approaches show good results. I remember that I had some jump up in the recent Microsoft Malware competition as well, and the winners used relatively simple models too.
", "rawMarkdown": "Thank you! Yes, it seems that in the competitions with such a massive shake up the simple approaches show good results. I remember that I had some jump up in the recent Microsoft Malware competition as well, and the winners used relatively simple models too." } ] }, { "id": 515021, "postDate": "2019-04-12T05:49:06.417Z", "content": "I don't have too much experience with time series classification tasks so I decided to go with a pretty simple approach. I used tsfresh library to extract a set of features from time series data. Also, the library helps to pick only relevant features among extracted using null-hypothesis testing framework. Then I trained a LightGBM classifier using the extracted features.
This kernel shows an example of this approach. The kernel gives 0.59 on private. However, if you extract more features, pick only the relevant ones, tune the classifier parameters and use k-fold validation, you'll probably get something around 0.66. (I am attaching a submission file that gives me that score on private).
Though this solution is far away from the top positions, it is very straightforward and could be implemented within a couple of hours. So it seems to be a good baseline solution for time series classification problems.
", "rawMarkdown": "I don't have too much experience with time series classification tasks so I decided to go with a pretty simple approach. I used `tsfresh` library to extract a set of features from time series data. Also, the library helps to pick only relevant features among extracted using null-hypothesis testing framework. Then I trained a LightGBM classifier using the extracted features. \n\n[This kernel shows](https://www.kaggle.com/purplejester/the-best-friend-of-an-alchemist?scriptVersionId=12730118) an example of this approach. The kernel gives `0.59` on private. However, if you extract more features, pick only the relevant ones, tune the classifier parameters and use k-fold validation, you'll probably get something around `0.66`. (I am attaching a submission file that gives me that score on private).\n\nThough this solution is far away from the top positions, it is very straightforward and could be implemented within a couple of hours. So it seems to be a good baseline solution for time series classification problems.", "votes": 1 } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T15:43:27Z", "competition_id": 13242, "competition_slug": "career-con-2019", "team_id": 2894626, "solution_url": "/competitions/career-con-2019/writeups/devforfu-the-92nd-place-and-378-positions-up-on-pr" }