Hi everyone!\nFirst of all I'm looking for teammates.\nMy current LB is -1.901. My basic concept was LGB on top of pre-engineered features, it was resulted in -1.718. Later I created MPNN using Keras, I tuned it for 3JHN type and expected that the result of other types will be improved correspondingly but in fact my Keras models gave good improvement for 3Jxx types and minimum improvement for 2JHH and 2JHN types and it couldn't make better result for 2JHC and 1Jxx types in comparison with LGB. So my -1.901 is a mix of not very good LGB for 1Jxx and you can say 2Jxx (because improvement by NN is minimal) and relatevely good NN for 3Jxx types.\nFor 3Jxx types I use 8-folds median and I found out that averaging give much better results that standalone model, so CV doesn't reflect the actual performance, my CV and estimation is as follows:\ntype - CV - actual performance estimation\n3JHN -2.59 -2.66\n3JHC -1.66 -1.70\n3JHH -2.38 -2.44\nThe experience of other teams shows that blending of different models can be much better than single models.\nSo if you are wondering it let's colaborate.
", "messages": [ { "id": 604848, "postDate": "2019-08-21T20:25:27.483Z", "content": "Hi everyone!\nFirst of all I'm looking for teammates.\nMy current LB is -1.901. My basic concept was LGB on top of pre-engineered features, it was resulted in -1.718. Later I created MPNN using Keras, I tuned it for 3JHN type and expected that the result of other types will be improved correspondingly but in fact my Keras models gave good improvement for 3Jxx types and minimum improvement for 2JHH and 2JHN types and it couldn't make better result for 2JHC and 1Jxx types in comparison with LGB. So my -1.901 is a mix of not very good LGB for 1Jxx and you can say 2Jxx (because improvement by NN is minimal) and relatevely good NN for 3Jxx types.\nFor 3Jxx types I use 8-folds median and I found out that averaging give much better results that standalone model, so CV doesn't reflect the actual performance, my CV and estimation is as follows:\ntype - CV - actual performance estimation\n3JHN -2.59 -2.66\n3JHC -1.66 -1.70\n3JHH -2.38 -2.44\nThe experience of other teams shows that blending of different models can be much better than single models.\nSo if you are wondering it let's colaborate.
", "rawMarkdown": "Hi everyone!\nFirst of all I'm looking for teammates.\nMy current LB is -1.901. My basic concept was LGB on top of pre-engineered features, it was resulted in -1.718. Later I created MPNN using Keras, I tuned it for 3JHN type and expected that the result of other types will be improved correspondingly but in fact my Keras models gave good improvement for 3Jxx types and minimum improvement for 2JHH and 2JHN types and it couldn't make better result for 2JHC and 1Jxx types in comparison with LGB. So my -1.901 is a mix of not very good LGB for 1Jxx and you can say 2Jxx (because improvement by NN is minimal) and relatevely good NN for 3Jxx types.\nFor 3Jxx types I use 8-folds median and I found out that averaging give much better results that standalone model, so CV doesn't reflect the actual performance, my CV and estimation is as follows:\ntype - CV - actual performance estimation\n3JHN -2.59 -2.66\n3JHC -1.66 -1.70\n3JHH -2.38 -2.44\nThe experience of other teams shows that blending of different models can be much better than single models.\nSo if you are wondering it let's colaborate." }, { "id": 604876, "postDate": "2019-08-21T20:48:39.430Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 604876, "author_name": "", "author_url": "", "post_date": "2019-08-21T20:48:39.430000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "604848": "Hi everyone!\nFirst of all I'm looking for teammates.\nMy current LB is -1.901. My basic concept was LGB on top of pre-engineered features, it was resulted in -1.718. Later I created MPNN using Keras, I tuned it for 3JHN type and expected that the result of other types will be improved correspondingly but in fact my Keras models gave good improvement for 3Jxx types and minimum improvement for 2JHH and 2JHN types and it couldn't make better result for 2JHC and 1Jxx types in comparison with LGB. So my -1.901 is a mix of not very good LGB for 1Jxx and you can say 2Jxx (because improvement by NN is minimal) and relatevely good NN for 3Jxx types.\nFor 3Jxx types I use 8-folds median and I found out that averaging give much better results that standalone model, so CV doesn't reflect the actual performance, my CV and estimation is as follows:\ntype - CV - actual performance estimation\n3JHN -2.59 -2.66\n3JHC -1.66 -1.70\n3JHH -2.38 -2.44\nThe experience of other teams shows that blending of different models can be much better than single models.\nSo if you are wondering it let's colaborate.", "604876": "" } }