Congratulations to the winners of this competition, and thanks Kaggle and Favorita for hosting it. My special thanks to @Ceshine and @Lingzhi for sharing their codes.
\n\nMy code was based on Ceshine's code, I added more features and played with it a lot trying to get better public LB. I had stuck at 0.513 public LB after several tries, and did not have enough time to try other solutions. Fortunately, I jumped 443 steps in private LB that made me smile today!
\n\nHere I share some points that I believe to be the reason for my improved performance on the test set:\n1- My model is an LGBM model with added features. Latest feature count was near 100 focusing more on daily, weekly and monthly features, plus categorical features focusing on items and stores.
\n\n2- Large train size: I increased the train data set to 19 weeks starting from 3/15/2017, and the validation set was the last possible week 7/26/2017. This gave a train set almost 3 times larger than public kernels, and most likely helped me to scape the 'curse of dimensionality' with the increased number of data points. I think there was still room to increase the train data size, specially the available data between 7/26 to 8/15 were not used properly. This could be better handled.
\n\n3- Forgetting factor (decay): I weighted early weeks of train data with a lower weight compared to the late weeks. This means that models are trained to be more dependent on the latest customer behavior rather than old patterns. \nI didn't make it to spend time on fine tuning the LGBM model, I wish I had!
", "messages": [ { "id": "269501", "postDate": "01/16/2018 23:39:47", "content": "Congratulations to the winners of this competition, and thanks Kaggle and Favorita for hosting it. My special thanks to @Ceshine and @Lingzhi for sharing their codes.
\n\nMy code was based on Ceshine's code, I added more features and played with it a lot trying to get better public LB. I had stuck at 0.513 public LB after several tries, and did not have enough time to try other solutions. Fortunately, I jumped 443 steps in private LB that made me smile today!
\n\nHere I share some points that I believe to be the reason for my improved performance on the test set:\n1- My model is an LGBM model with added features. Latest feature count was near 100 focusing more on daily, weekly and monthly features, plus categorical features focusing on items and stores.
\n\n2- Large train size: I increased the train data set to 19 weeks starting from 3/15/2017, and the validation set was the last possible week 7/26/2017. This gave a train set almost 3 times larger than public kernels, and most likely helped me to scape the 'curse of dimensionality' with the increased number of data points. I think there was still room to increase the train data size, specially the available data between 7/26 to 8/15 were not used properly. This could be better handled.
\n\n3- Forgetting factor (decay): I weighted early weeks of train data with a lower weight compared to the late weeks. This means that models are trained to be more dependent on the latest customer behavior rather than old patterns. \nI didn't make it to spend time on fine tuning the LGBM model, I wish I had!
", "rawMarkdown": "Congratulations to the winners of this competition, and thanks Kaggle and Favorita for hosting it. My special thanks to @Ceshine and @Lingzhi for sharing their codes.\n\nMy code was based on Ceshine's code, I added more features and played with it a lot trying to get better public LB. I had stuck at 0.513 public LB after several tries, and did not have enough time to try other solutions. Fortunately, I jumped 443 steps in private LB that made me smile today! \n\nHere I share some points that I believe to be the reason for my improved performance on the test set:\n1- My model is an LGBM model with added features. Latest feature count was near 100 focusing more on daily, weekly and monthly features, plus categorical features focusing on items and stores.\n\n2- Large train size: I increased the train data set to 19 weeks starting from 3/15/2017, and the validation set was the last possible week 7/26/2017. This gave a train set almost 3 times larger than public kernels, and most likely helped me to scape the 'curse of dimensionality' with the increased number of data points. I think there was still room to increase the train data size, specially the available data between 7/26 to 8/15 were not used properly. This could be better handled.\n\n3- Forgetting factor (decay): I weighted early weeks of train data with a lower weight compared to the late weeks. This means that models are trained to be more dependent on the latest customer behavior rather than old patterns. \nI didn't make it to spend time on fine tuning the LGBM model, I wish I had!", "votes": null }, { "id": "269533", "postDate": "01/17/2018 01:02:38", "content": "Thanks for sharing. How helpful was the decay factor? Seems like an idea I should try next time!
", "rawMarkdown": "Thanks for sharing. How helpful was the decay factor? Seems like an idea I should try next time!", "votes": null }, { "id": "269557", "postDate": "01/17/2018 01:46:54", "content": "According to 1st kernel, the decay mean get highest feature importance score. Worth a try.
", "rawMarkdown": "According to 1st kernel, the decay mean get highest feature importance score. Worth a try.", "votes": null }, { "id": "270194", "postDate": "01/17/2018 22:29:01", "content": "I don't have a separate submission with decay as the only difference, so unfortunately I cannot give quantitative answer to your question. However, locally I observed an improvement in the validation set (~0.005) when I applied that. I had a gut feeling that it's useful.
", "rawMarkdown": "I don't have a separate submission with decay as the only difference, so unfortunately I cannot give quantitative answer to your question. However, locally I observed an improvement in the validation set (~0.005) when I applied that. I had a gut feeling that it's useful.", "votes": null }, { "id": "270417", "postDate": "01/18/2018 07:37:42", "content": "Thanks.
\n\n\n\n\nI had a gut feeling that it's useful.
\n
I have the same gut feeling.
", "rawMarkdown": "Thanks.\n\n> I had a gut feeling that it's useful.\n\nI have the same gut feeling.", "votes": null } ], "comments": [ { "id": 269533, "author_name": "cpmpml", "author_url": "", "post_date": "01/17/2018 01:02:38", "content": "Thanks for sharing. How helpful was the decay factor? Seems like an idea I should try next time!
", "votes": null, "replies": [ { "id": 269557, "author_name": "yedthon", "author_url": "", "post_date": "01/17/2018 01:46:54", "content": "According to 1st kernel, the decay mean get highest feature importance score. Worth a try.
", "votes": null, "replies": [] }, { "id": 270194, "author_name": "faridtajad", "author_url": "", "post_date": "01/17/2018 22:29:01", "content": "I don't have a separate submission with decay as the only difference, so unfortunately I cannot give quantitative answer to your question. However, locally I observed an improvement in the validation set (~0.005) when I applied that. I had a gut feeling that it's useful.
", "votes": null, "replies": [] }, { "id": 270417, "author_name": "cpmpml", "author_url": "", "post_date": "01/18/2018 07:37:42", "content": "Thanks.
\n\n\n\n\nI had a gut feeling that it's useful.
\n
I have the same gut feeling.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "269501": "Congratulations to the winners of this competition, and thanks Kaggle and Favorita for hosting it. My special thanks to @Ceshine and @Lingzhi for sharing their codes.\n\nMy code was based on Ceshine's code, I added more features and played with it a lot trying to get better public LB. I had stuck at 0.513 public LB after several tries, and did not have enough time to try other solutions. Fortunately, I jumped 443 steps in private LB that made me smile today! \n\nHere I share some points that I believe to be the reason for my improved performance on the test set:\n1- My model is an LGBM model with added features. Latest feature count was near 100 focusing more on daily, weekly and monthly features, plus categorical features focusing on items and stores.\n\n2- Large train size: I increased the train data set to 19 weeks starting from 3/15/2017, and the validation set was the last possible week 7/26/2017. This gave a train set almost 3 times larger than public kernels, and most likely helped me to scape the 'curse of dimensionality' with the increased number of data points. I think there was still room to increase the train data size, specially the available data between 7/26 to 8/15 were not used properly. This could be better handled.\n\n3- Forgetting factor (decay): I weighted early weeks of train data with a lower weight compared to the late weeks. This means that models are trained to be more dependent on the latest customer behavior rather than old patterns. \nI didn't make it to spend time on fine tuning the LGBM model, I wish I had!", "269533": "Thanks for sharing. How helpful was the decay factor? Seems like an idea I should try next time!", "269557": "According to 1st kernel, the decay mean get highest feature importance score. Worth a try.", "270194": "I don't have a separate submission with decay as the only difference, so unfortunately I cannot give quantitative answer to your question. However, locally I observed an improvement in the validation set (~0.005) when I applied that. I had a gut feeling that it's useful.", "270417": "Thanks.\n\n> I had a gut feeling that it's useful.\n\nI have the same gut feeling." }, "source": "meta" }