{ "id": 596305, "title": "Any Success or Lift from Having a Phase Detection Feature", "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/596305", "author_name": "", "post_date": "2025-08-02T19:03:39.653058100Z", "votes": 2, "comment_count": 4, "views": 0, "content": "

Has anyone tried out using a seq2seq model to determine whether a timestamp is a transition or a gesture and passing it in a feature? My intuition tells me that passing that info to a downstream classification model should help it focus on the portion that is a gesture. On top of that, the transitions might be just giving garbage noise for a classification model to parse through, and since we don't have a bunch of example relatively speaking, it should give it some help.

\n

….but then again I haven't seen it in too many high ranking notebooks so curious if anyone has tried it yet.

", "messages": [ { "id": "3262035", "postDate": "08/02/2025 19:03:39", "content": "

Has anyone tried out using a seq2seq model to determine whether a timestamp is a transition or a gesture and passing it in a feature? My intuition tells me that passing that info to a downstream classification model should help it focus on the portion that is a gesture. On top of that, the transitions might be just giving garbage noise for a classification model to parse through, and since we don't have a bunch of example relatively speaking, it should give it some help.

\n

….but then again I haven't seen it in too many high ranking notebooks so curious if anyone has tried it yet.

", "rawMarkdown": "Has anyone tried out using a seq2seq model to determine whether a timestamp is a transition or a gesture and passing it in a feature? My intuition tells me that passing that info to a downstream classification model should help it focus on the portion that is a gesture. On top of that, the transitions might be just giving garbage noise for a classification model to parse through, and since we don't have a bunch of example relatively speaking, it should give it some help.\n\n....but then again I haven't seen it in too many high ranking notebooks so curious if anyone has tried it yet.", "votes": null }, { "id": "3262048", "postDate": "08/02/2025 19:17:39", "content": "

you can try passing phase info as feature to see if its worth going that path

", "rawMarkdown": "you can try passing phase info as feature to see if its worth going that path", "votes": null }, { "id": "3262325", "postDate": "08/03/2025 12:09:42", "content": "

I used phase as an auxiliary label for seq2seq training, i.e., in a multi-task learning setup, but it didn’t lead to any improvement in my cv score.

", "rawMarkdown": "I used phase as an auxiliary label for seq2seq training, i.e., in a multi-task learning setup, but it didn’t lead to any improvement in my cv score.", "votes": null }, { "id": "3262538", "postDate": "08/03/2025 19:56:55", "content": "

That's plan at the moment. Worst case, I brush up on my seq2seq training.

", "rawMarkdown": "That's plan at the moment. Worst case, I brush up on my seq2seq training.", "votes": null }, { "id": "3262881", "postDate": "08/04/2025 12:07:29", "content": "

I have a phase head and I use masked bce loss (0 for transition, 1 for gesture, mask value for padded indices). It improves my score little bit but I'm not sure if it's significant enough.

", "rawMarkdown": "I have a phase head and I use masked bce loss (0 for transition, 1 for gesture, mask value for padded indices). It improves my score little bit but I'm not sure if it's significant enough.", "votes": null } ], "comments": [ { "id": 3262048, "author_name": "celiker", "author_url": "", "post_date": "08/02/2025 19:17:39", "content": "

you can try passing phase info as feature to see if its worth going that path

", "votes": null, "replies": [ { "id": 3262538, "author_name": "msthil", "author_url": "", "post_date": "08/03/2025 19:56:55", "content": "

That's plan at the moment. Worst case, I brush up on my seq2seq training.

", "votes": null, "replies": [] } ] }, { "id": 3262325, "author_name": "pshuaibi", "author_url": "", "post_date": "08/03/2025 12:09:42", "content": "

I used phase as an auxiliary label for seq2seq training, i.e., in a multi-task learning setup, but it didn’t lead to any improvement in my cv score.

", "votes": null, "replies": [] }, { "id": 3262881, "author_name": "gunesevitan", "author_url": "", "post_date": "08/04/2025 12:07:29", "content": "

I have a phase head and I use masked bce loss (0 for transition, 1 for gesture, mask value for padded indices). It improves my score little bit but I'm not sure if it's significant enough.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3262035": "Has anyone tried out using a seq2seq model to determine whether a timestamp is a transition or a gesture and passing it in a feature? My intuition tells me that passing that info to a downstream classification model should help it focus on the portion that is a gesture. On top of that, the transitions might be just giving garbage noise for a classification model to parse through, and since we don't have a bunch of example relatively speaking, it should give it some help.\n\n....but then again I haven't seen it in too many high ranking notebooks so curious if anyone has tried it yet.", "3262048": "you can try passing phase info as feature to see if its worth going that path", "3262325": "I used phase as an auxiliary label for seq2seq training, i.e., in a multi-task learning setup, but it didn’t lead to any improvement in my cv score.", "3262538": "That's plan at the moment. Worst case, I brush up on my seq2seq training.", "3262881": "I have a phase head and I use masked bce loss (0 for transition, 1 for gesture, mask value for padded indices). It improves my score little bit but I'm not sure if it's significant enough." }, "source": "meta" }