Congrats to all the winners!
\nAnd I must apologize Kaggle community for cheating. This competition was so hard that I made a mistake.
\nHere, I want to share my part of our team's solution. (My teammate higepon's part is here)
My wifi model is CNN. Divide ±10 seconds into 5 sections and create (1, 5, rssi) images as input feature.
\nI also created sensor RNN model. This may be an attempt that many participants have not made.
\nCongrats to all the winners!
\nAnd I must apologize Kaggle community for cheating. This competition was so hard that I made a mistake.
\nHere, I want to share my part of our team's solution. (My teammate higepon's part is here)
My wifi model is CNN. Divide ±10 seconds into 5 sections and create (1, 5, rssi) images as input feature.
\nI also created sensor RNN model. This may be an attempt that many participants have not made.
\nHey man, that's not cheating. You explored something, shared it on the forum, and I believe you didn't use it finally. Some kaggle competitions do allow hand labeling. it is a case-by-case thing.
", "rawMarkdown": "Hey man, that's not cheating. You explored something, shared it on the forum, and I believe you didn't use it finally. Some kaggle competitions do allow hand labeling. it is a case-by-case thing.", "votes": null }, { "id": "1313717", "postDate": "05/18/2021 17:11:16", "content": "Thank you for your comment.
\nThe criteria for OK/NG is difficult for Kaggle beginners. So I would not use hand labeling in Kaggle anymore.
\nAnd, your comment \"the most magical kernel\" was very impressive for me. Thank you very much.
For even experienced kagglers, this case is really difficult to judge, because it's unclear whether hand-labelled possible missing grids are hand-labelling for test set or training set. We all know your team did NOT cheat, and your team has been always very fair.
", "rawMarkdown": "For even experienced kagglers, this case is really difficult to judge, because it's unclear whether hand-labelled possible missing grids are hand-labelling for test set or training set. We all know your team did NOT cheat, and your team has been always very fair.", "votes": null }, { "id": "1313819", "postDate": "05/18/2021 18:23:18", "content": "Thanks akio, I sincerely respect your great achievement and this fair and open attitude. Without your great cost minimizaiton notebook, our team could never have been 2nd. I believe you are the person who is most respected by the participants of this competition.
", "rawMarkdown": "Thanks akio, I sincerely respect your great achievement and this fair and open attitude. Without your great cost minimizaiton notebook, our team could never have been 2nd. I believe you are the person who is most respected by the participants of this competition.", "votes": null } ], "comments": [ { "id": 1313615, "author_name": "jiweiliu", "author_url": "", "post_date": "05/18/2021 16:15:08", "content": "Hey man, that's not cheating. You explored something, shared it on the forum, and I believe you didn't use it finally. Some kaggle competitions do allow hand labeling. it is a case-by-case thing.
", "votes": null, "replies": [ { "id": 1313717, "author_name": "saitodevel01", "author_url": "", "post_date": "05/18/2021 17:11:16", "content": "Thank you for your comment.
\nThe criteria for OK/NG is difficult for Kaggle beginners. So I would not use hand labeling in Kaggle anymore.
\nAnd, your comment \"the most magical kernel\" was very impressive for me. Thank you very much.
For even experienced kagglers, this case is really difficult to judge, because it's unclear whether hand-labelled possible missing grids are hand-labelling for test set or training set. We all know your team did NOT cheat, and your team has been always very fair.
", "votes": null, "replies": [] } ] }, { "id": 1313819, "author_name": "mamasinkgs", "author_url": "", "post_date": "05/18/2021 18:23:18", "content": "Thanks akio, I sincerely respect your great achievement and this fair and open attitude. Without your great cost minimizaiton notebook, our team could never have been 2nd. I believe you are the person who is most respected by the participants of this competition.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1313372": "Congrats to all the winners!\nAnd I must apologize Kaggle community for [cheating](https://www.kaggle.com/c/indoor-location-navigation/discussion/235813). This competition was so hard that I made a mistake. \nHere, I want to share my part of our team's solution. (My teammate higepon's part is [here](https://www.kaggle.com/c/indoor-location-navigation/discussion/239919))\n\n## Wifi models (floor and absolute position)\n\nMy wifi model is CNN. Divide ±10 seconds into 5 sections and create (1, 5, rssi) images as input feature.\n\n+ Adding ibeacon's rssi in input features.\n+ Site specific models.\n+ Use compute_step_positions of host code for data augmentation.\n+ Use StratifiedGroupKFold (label=floor, group=path) for both floor and position models.\n+ Error estimation based on variance between different models (including public notebooks).\n+ My wifi model's performance is poor compared to the models discussed [here](https://www.kaggle.com/c/indoor-location-navigation/discussion/232995). (CV/LB : 7.1 to 7.3) But relatively small gap between CV, public LB and private LB.\n+ Thank you very much @higepon for this part.\n\n## Sensor models (relative position)\n\nI also created sensor RNN model. This may be an attempt that many participants have not made.\n\n+ Apply anti-aliasing filter to sensor data and resample them to sampling rate 0.1 seconds to reduce data.\n+ Try LSTM and GRU and ensemble them.\n+ Error estimation based on ensemble variance.\n+ Calibrate scale and rotation for each device using device leak. This boosts public/private LB 0.03m.\n+ Adding uncalibrated data didn't work for my model.\n\n## Post-processing\n\n+ My cost minimization notebook is equivalent to Kalman smoother (See [this notebook](https://www.kaggle.com/saitodevel01/background-ideas-of-cost-minimization-notebook) for detail).\n+ Parameters of cost function are determined by error estimation of wifi and sensor models.\n+ Heuristics that gradually replace wifi predictions with nearest neighbor grid points.\n+ Use additional grid points generated by [this code](https://www.kaggle.com/saitodevel01/generate-extra-grid-points).\n+ [I published our final submission code.](https://www.kaggle.com/saitodevel01/11-pseudo-labeling-from-lb-2-586-with-retry)", "1313615": "Hey man, that's not cheating. You explored something, shared it on the forum, and I believe you didn't use it finally. Some kaggle competitions do allow hand labeling. it is a case-by-case thing.", "1313717": "Thank you for your comment.\nThe criteria for OK/NG is difficult for Kaggle beginners. So I would not use hand labeling in Kaggle anymore.\nAnd, your comment \"the most magical kernel\" was very impressive for me. Thank you very much.", "1313807": "For even experienced kagglers, this case is really difficult to judge, because it's unclear whether hand-labelled possible missing grids are hand-labelling for test set or training set. We all know your team did NOT cheat, and your team has been always very fair.", "1313819": "Thanks akio, I sincerely respect your great achievement and this fair and open attitude. Without your great cost minimizaiton notebook, our team could never have been 2nd. I believe you are the person who is most respected by the participants of this competition." }, "source": "meta" }