To reach the top positions what strategy they are using.
\na) RNN + floor 99% + cost minimization + snap_to_grid
\nb) Ensemble (lgb + rf + nn +rnn) + floor 99% + cost minimization + snap_to_grid
I suppose that the big difference is knowing how to get the maximum information out of the data as well as the associated feature engineering techniques.
\nWhat do you think?π€
", "messages": [ { "id": "1252497", "postDate": "03/25/2021 18:10:23", "content": "To reach the top positions what strategy they are using.
\na) RNN + floor 99% + cost minimization + snap_to_grid
\nb) Ensemble (lgb + rf + nn +rnn) + floor 99% + cost minimization + snap_to_grid
I suppose that the big difference is knowing how to get the maximum information out of the data as well as the associated feature engineering techniques.
\nWhat do you think?π€
", "rawMarkdown": "To reach the top positions what strategy they are using.\n\na) RNN + floor 99% + cost minimization + snap_to_grid\nb) Ensemble (lgb + rf + nn +rnn) + floor 99% + cost minimization + snap_to_grid\n\nI suppose that the big difference is knowing how to get the maximum information out of the data as well as the associated feature engineering techniques.\n\nWhat do you think?π€", "votes": null }, { "id": "1254330", "postDate": "03/27/2021 14:39:42", "content": "Here is what I recommend:
\n1) Download and generate multiple different features/datasets.
\n2) Incorporate the 99% floor preds as a feature (I have a discussion about this here: https://www.kaggle.com/c/indoor-location-navigation/discussion/228887)
\n3) Ensemble multiple rnns that each use a different dataset and maybe include a lgb in your ensemble too
\n4) Apply cost minimization followed by snap to grid
I donβt know what the people at the very top are doing but this has gotten me top 30 and a decent margin above the top published score.
", "rawMarkdown": "Here is what I recommend: \n1) Download and generate multiple different features/datasets.\n2) Incorporate the 99% floor preds as a feature (I have a discussion about this here: [https://www.kaggle.com/c/indoor-location-navigation/discussion/228887](url))\n3) Ensemble multiple rnns that each use a different dataset and maybe include a lgb in your ensemble too\n4) Apply cost minimization followed by snap to grid\n\nI donβt know what the people at the very top are doing but this has gotten me top 30 and a decent margin above the top published score.", "votes": null }, { "id": "1254681", "postDate": "03/28/2021 01:08:03", "content": "Not sure what you call top positions but here's what i'm doing:
\nI divided the indoor position prediction task into two quite distinct parts:
\n1) the usage of WIFI features to get an approximate location of pedestrians. I used time series form modelling to avoid that the obtained positions follow inconsistent trajectories. More specifically, custom RNNs which allows the model to capture longer sequences and therefore acts similarly as a \"time-regularized transformers\". I used an ensemble of multiple models with different seeds which lead to a huge gain in my LB ranking. I also tried plenty of stuff related to my thesis (multi scale temporal modelling) but doesn't really changes anything.
\n2) the usage of other data provided in the dataset to smooth the estimation of trajectories and the application of all the post processing methods proposed by way more knowledgable kagglers than me since it's not my domain.
\nWhat I can tell you however, is that the floor 99% acc submission is robust in Public but not so much reliable in Private LB, so I also specifically worked on increasing my floor prediction even tho it doesn't really show up on Public LBβ¦
", "rawMarkdown": "Not sure what you call top positions but here's what i'm doing:\n\nI divided the indoor position prediction task into two quite distinct parts:\n\n 1) the usage of WIFI features to get an approximate location of pedestrians. I used time series form modelling to avoid that the obtained positions follow inconsistent trajectories. More specifically, custom RNNs which allows the model to capture longer sequences and therefore acts similarly as a \"time-regularized transformers\". I used an ensemble of multiple models with different seeds which lead to a huge gain in my LB ranking. I also tried plenty of stuff related to my thesis (multi scale temporal modelling) but doesn't really changes anything.\n\n2) the usage of other data provided in the dataset to smooth the estimation of trajectories and the application of all the post processing methods proposed by way more knowledgable kagglers than me since it's not my domain.\n\n\nWhat I can tell you however, is that the floor 99% acc submission is robust in Public but not so much reliable in Private LB, so I also specifically worked on increasing my floor prediction even tho it doesn't really show up on Public LB...", "votes": null }, { "id": "1255037", "postDate": "03/28/2021 11:10:51", "content": "Ok, of course you have cleared the way for me.
\nI wonder how to use the variable \"timestamp\" in the rnn.
\nThanks a lot @ravishah1 π
Quite interesting Joseph, I think I don't know exactly how to set up the infrastructure to use the data as time series, I have to investigate it. ( Is there a public notebook to understand the approach you are commenting on?)
\nOn the other hand, what you say about the predictions of the plants makes sense, how did you discover that?
Thank you very much for the feedback.π
", "rawMarkdown": "Quite interesting Joseph, I think I don't know exactly how to set up the infrastructure to use the data as time series, I have to investigate it. ( Is there a public notebook to understand the approach you are commenting on?)\nOn the other hand, what you say about the predictions of the plants makes sense, how did you discover that?\n\nThank you very much for the feedback.π", "votes": null } ], "comments": [ { "id": 1254330, "author_name": "ravishah1", "author_url": "", "post_date": "03/27/2021 14:39:42", "content": "Here is what I recommend:
\n1) Download and generate multiple different features/datasets.
\n2) Incorporate the 99% floor preds as a feature (I have a discussion about this here: https://www.kaggle.com/c/indoor-location-navigation/discussion/228887)
\n3) Ensemble multiple rnns that each use a different dataset and maybe include a lgb in your ensemble too
\n4) Apply cost minimization followed by snap to grid
I donβt know what the people at the very top are doing but this has gotten me top 30 and a decent margin above the top published score.
", "votes": null, "replies": [ { "id": 1255037, "author_name": "trasibulo", "author_url": "", "post_date": "03/28/2021 11:10:51", "content": "Ok, of course you have cleared the way for me.
\nI wonder how to use the variable \"timestamp\" in the rnn.
\nThanks a lot @ravishah1 π
Not sure what you call top positions but here's what i'm doing:
\nI divided the indoor position prediction task into two quite distinct parts:
\n1) the usage of WIFI features to get an approximate location of pedestrians. I used time series form modelling to avoid that the obtained positions follow inconsistent trajectories. More specifically, custom RNNs which allows the model to capture longer sequences and therefore acts similarly as a \"time-regularized transformers\". I used an ensemble of multiple models with different seeds which lead to a huge gain in my LB ranking. I also tried plenty of stuff related to my thesis (multi scale temporal modelling) but doesn't really changes anything.
\n2) the usage of other data provided in the dataset to smooth the estimation of trajectories and the application of all the post processing methods proposed by way more knowledgable kagglers than me since it's not my domain.
\nWhat I can tell you however, is that the floor 99% acc submission is robust in Public but not so much reliable in Private LB, so I also specifically worked on increasing my floor prediction even tho it doesn't really show up on Public LBβ¦
", "votes": null, "replies": [ { "id": 1255045, "author_name": "trasibulo", "author_url": "", "post_date": "03/28/2021 11:14:59", "content": "Quite interesting Joseph, I think I don't know exactly how to set up the infrastructure to use the data as time series, I have to investigate it. ( Is there a public notebook to understand the approach you are commenting on?)
\nOn the other hand, what you say about the predictions of the plants makes sense, how did you discover that?
Thank you very much for the feedback.π
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1252497": "To reach the top positions what strategy they are using.\n\na) RNN + floor 99% + cost minimization + snap_to_grid\nb) Ensemble (lgb + rf + nn +rnn) + floor 99% + cost minimization + snap_to_grid\n\nI suppose that the big difference is knowing how to get the maximum information out of the data as well as the associated feature engineering techniques.\n\nWhat do you think?π€", "1254330": "Here is what I recommend: \n1) Download and generate multiple different features/datasets.\n2) Incorporate the 99% floor preds as a feature (I have a discussion about this here: [https://www.kaggle.com/c/indoor-location-navigation/discussion/228887](url))\n3) Ensemble multiple rnns that each use a different dataset and maybe include a lgb in your ensemble too\n4) Apply cost minimization followed by snap to grid\n\nI donβt know what the people at the very top are doing but this has gotten me top 30 and a decent margin above the top published score.", "1254681": "Not sure what you call top positions but here's what i'm doing:\n\nI divided the indoor position prediction task into two quite distinct parts:\n\n 1) the usage of WIFI features to get an approximate location of pedestrians. I used time series form modelling to avoid that the obtained positions follow inconsistent trajectories. More specifically, custom RNNs which allows the model to capture longer sequences and therefore acts similarly as a \"time-regularized transformers\". I used an ensemble of multiple models with different seeds which lead to a huge gain in my LB ranking. I also tried plenty of stuff related to my thesis (multi scale temporal modelling) but doesn't really changes anything.\n\n2) the usage of other data provided in the dataset to smooth the estimation of trajectories and the application of all the post processing methods proposed by way more knowledgable kagglers than me since it's not my domain.\n\n\nWhat I can tell you however, is that the floor 99% acc submission is robust in Public but not so much reliable in Private LB, so I also specifically worked on increasing my floor prediction even tho it doesn't really show up on Public LB...", "1255037": "Ok, of course you have cleared the way for me.\nI wonder how to use the variable \"timestamp\" in the rnn.\nThanks a lot @ravishah1 π", "1255045": "Quite interesting Joseph, I think I don't know exactly how to set up the infrastructure to use the data as time series, I have to investigate it. ( Is there a public notebook to understand the approach you are commenting on?)\nOn the other hand, what you say about the predictions of the plants makes sense, how did you discover that?\n\nThank you very much for the feedback.π" }, "source": "meta" }