![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F443651%2Fcd6617a65b5db23e9ed2389e11e96874%2Fdistribution.png?generation=1674766330748997&alt=media\")
I have published a notebook in which I analyse the similarities in public & private leaderboard ranks of pairs of users. A similar suggestion was made by @philippsinger here. The notebook can be vastly improved, but I feel that it could be a good starting point for others to continue with, hence why I am publishing it.
\nThe goal of this notebook is to generate some data-driven evidence of potential private sharing, whilst waiting on Kaggle to release their cheating detection competition :).
\nSome potential improvement points:
\nBelow is a plot of mean absolute difference of public ranks + mean absolute difference of private ranks between users & a vertical line of two users mentioned in thread within this forum. Note that I do not claim this to be hard evidence whatsoever.
\n
The notebook can be found here
", "messages": [ { "id": "2116956", "postDate": "01/26/2023 20:52:17", "content": "I have published a notebook in which I analyse the similarities in public & private leaderboard ranks of pairs of users. A similar suggestion was made by @philippsinger here. The notebook can be vastly improved, but I feel that it could be a good starting point for others to continue with, hence why I am publishing it.
\nThe goal of this notebook is to generate some data-driven evidence of potential private sharing, whilst waiting on Kaggle to release their cheating detection competition :).
\nSome potential improvement points:
\nBelow is a plot of mean absolute difference of public ranks + mean absolute difference of private ranks between users & a vertical line of two users mentioned in thread within this forum. Note that I do not claim this to be hard evidence whatsoever.
\n
The notebook can be found here
", "rawMarkdown": "I have published a notebook in which I analyse the similarities in public & private leaderboard ranks of pairs of users. A similar suggestion was made by @philippsinger [here](https://www.kaggle.com/competitions/otto-recommender-system/discussion/381360#2116335). The notebook can be vastly improved, but I feel that it could be a good starting point for others to continue with, hence why I am publishing it.\n\nThe goal of this notebook is to generate some data-driven evidence of potential private sharing, whilst waiting on Kaggle to release their cheating detection competition :).\n\nSome potential improvement points:\n* Improving the pairwise distance / similarity metric (example: measure difference in shakeups)\n* Graph analysis\n* Analysis using multiple metrics\n* Running on more data (currently strict filtering has been applied, such as e.g. only looking at masters & GMs)\n* etc.\n\nBelow is a plot of mean absolute difference of public ranks + mean absolute difference of private ranks between users & a vertical line of two users mentioned in thread within this forum. **Note that I do not claim this to be hard evidence whatsoever.**\n\n\n\nThe notebook can be found [here](https://www.kaggle.com/code/group16/lb-rank-analysis/notebook)", "votes": null }, { "id": "2117304", "postDate": "01/27/2023 07:19:40", "content": "Imo these two people are probably not innocent, but this is not a good analysis that in fact can be their proof that they are not related!
\nSince kaggle master/grandmaster 1. join lots of competition 2. are good at it, there will be higher chance to get a high correlation just by coincident.
\nIn face these two are not even having the highest correlation among other competitors (have to state here strongly that any people that are high on this list are not cheating)
\n ![\"\"](\"https://i.imgur.com/i1dadco.png\")
\n ![\"\"](\"https://i.imgur.com/EMgd6IQ.png\")
For me the correlation between submission is probably only thing that can be definitive proof.
", "rawMarkdown": "Imo these two people are probably not innocent, but this is not a good analysis that in fact can be their proof that they are not related! \nSince kaggle master/grandmaster 1. join lots of competition 2. are good at it, there will be higher chance to get a high correlation just by coincident. \nIn face these two are not even having the highest correlation among other competitors (**have to state here strongly that any people that are high on this list are not cheating**)\n \n \n\nFor me the correlation between submission is probably only thing that can be definitive proof.", "votes": null }, { "id": "2117341", "postDate": "01/27/2023 07:48:52", "content": "There is one important thing missing here. Those two always appear in gold zone during last weeks of competitions. I don't know how we can incorporate this information though.
", "rawMarkdown": "There is one important thing missing here. Those two always appear in gold zone during last weeks of competitions. I don't know how we can incorporate this information though.", "votes": null }, { "id": "2117382", "postDate": "01/27/2023 08:34:04", "content": "I think there is information on every submission made by a team in the meta-data!
", "rawMarkdown": "I think there is information on every submission made by a team in the meta-data!", "votes": null } ], "comments": [ { "id": 2117304, "author_name": "s903124", "author_url": "", "post_date": "01/27/2023 07:19:40", "content": "Imo these two people are probably not innocent, but this is not a good analysis that in fact can be their proof that they are not related!
\nSince kaggle master/grandmaster 1. join lots of competition 2. are good at it, there will be higher chance to get a high correlation just by coincident.
\nIn face these two are not even having the highest correlation among other competitors (have to state here strongly that any people that are high on this list are not cheating)
\n
\n
For me the correlation between submission is probably only thing that can be definitive proof.
", "votes": null, "replies": [] }, { "id": 2117341, "author_name": "gunesevitan", "author_url": "", "post_date": "01/27/2023 07:48:52", "content": "There is one important thing missing here. Those two always appear in gold zone during last weeks of competitions. I don't know how we can incorporate this information though.
", "votes": null, "replies": [ { "id": 2117382, "author_name": "group16", "author_url": "", "post_date": "01/27/2023 08:34:04", "content": "I think there is information on every submission made by a team in the meta-data!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2116956": "I have published a notebook in which I analyse the similarities in public & private leaderboard ranks of pairs of users. A similar suggestion was made by @philippsinger [here](https://www.kaggle.com/competitions/otto-recommender-system/discussion/381360#2116335). The notebook can be vastly improved, but I feel that it could be a good starting point for others to continue with, hence why I am publishing it.\n\nThe goal of this notebook is to generate some data-driven evidence of potential private sharing, whilst waiting on Kaggle to release their cheating detection competition :).\n\nSome potential improvement points:\n* Improving the pairwise distance / similarity metric (example: measure difference in shakeups)\n* Graph analysis\n* Analysis using multiple metrics\n* Running on more data (currently strict filtering has been applied, such as e.g. only looking at masters & GMs)\n* etc.\n\nBelow is a plot of mean absolute difference of public ranks + mean absolute difference of private ranks between users & a vertical line of two users mentioned in thread within this forum. **Note that I do not claim this to be hard evidence whatsoever.**\n\n\n\nThe notebook can be found [here](https://www.kaggle.com/code/group16/lb-rank-analysis/notebook)", "2117304": "Imo these two people are probably not innocent, but this is not a good analysis that in fact can be their proof that they are not related! \nSince kaggle master/grandmaster 1. join lots of competition 2. are good at it, there will be higher chance to get a high correlation just by coincident. \nIn face these two are not even having the highest correlation among other competitors (**have to state here strongly that any people that are high on this list are not cheating**)\n \n \n\nFor me the correlation between submission is probably only thing that can be definitive proof.", "2117341": "There is one important thing missing here. Those two always appear in gold zone during last weeks of competitions. I don't know how we can incorporate this information though.", "2117382": "I think there is information on every submission made by a team in the meta-data!" }, "source": "meta" }