![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F24422821%2Fbf183ca3ff252d23b391c4bb28897849%2FScreenshot%202025-06-03%20004807.png?generation=1748900961851735&alt=media\")
First of all, I want to thank the organizers for this wonderful and absolutely fruitful competition, learned a lot from it!
\nI used the original dataset provided in the competition, with one modification: I removed the top 100 training samples that had the most missing values. I then imputed the remaining missing values using the mean of each column from the training set.
\nI tried several approaches other than this one, all gave me worse results on public LB. Another thing that is worth mentioning is that I tried another model architecture (transformer encoder) but without dropping samples from the original dataset, that time I imputed the data with KNN. Results on public LB were no good compared to what I was getting from other approaches, but on private LB this submission was the best among others. It would've gotten me the 3rd place if I had chosen it :)
\n
To model the complex and spatially structured geological data, I designed a hybrid architecture that combines the global modeling power of Transformer encoders with the spatial precision of a UNet-style convolutional decoder.
\nI did not use any ensemble techniques—this result was achieved using a single model. You can find the full training notebook here:
\nđź”—https://www.kaggle.com/code/abdulkareemomer/geology-forecast-challange
", "comments": [], "comment_count": 0, "messages": [ { "id": 3215944, "postDate": "2025-06-02T22:19:32.317Z", "content": "First of all, I want to thank the organizers for this wonderful and absolutely fruitful competition, learned a lot from it!
\nI used the original dataset provided in the competition, with one modification: I removed the top 100 training samples that had the most missing values. I then imputed the remaining missing values using the mean of each column from the training set.
\nI tried several approaches other than this one, all gave me worse results on public LB. Another thing that is worth mentioning is that I tried another model architecture (transformer encoder) but without dropping samples from the original dataset, that time I imputed the data with KNN. Results on public LB were no good compared to what I was getting from other approaches, but on private LB this submission was the best among others. It would've gotten me the 3rd place if I had chosen it :)
\n
To model the complex and spatially structured geological data, I designed a hybrid architecture that combines the global modeling power of Transformer encoders with the spatial precision of a UNet-style convolutional decoder.
\nI did not use any ensemble techniques—this result was achieved using a single model. You can find the full training notebook here:
\nđź”—https://www.kaggle.com/code/abdulkareemomer/geology-forecast-challange
", "rawMarkdown": "First of all, I want to thank the organizers for this wonderful and absolutely fruitful competition, learned a lot from it!\n### DATA\nI used the original dataset provided in the competition, with one modification: I removed the top 100 training samples that had the most missing values. I then imputed the remaining missing values using the mean of each column from the training set.\n\nI tried several approaches other than this one, all gave me worse results on public LB. Another thing that is worth mentioning is that I tried another model architecture (transformer encoder) but without dropping samples from the original dataset, that time I imputed the data with KNN. Results on public LB were no good compared to what I was getting from other approaches, but on private LB this submission was the best among others. It would've gotten me the 3rd place if I had chosen it :)\n\n\n\n### MODEL AND TRAINING \nTo model the complex and spatially structured geological data, I designed a hybrid architecture that combines the global modeling power of Transformer encoders with the spatial precision of a UNet-style convolutional decoder.\n\nI did not use any ensemble techniques—this result was achieved using a single model. You can find the full training notebook here:\n\n🔗https://www.kaggle.com/code/abdulkareemomer/geology-forecast-challange", "votes": 3 } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T12:42:37Z", "competition_id": 95697, "competition_slug": "geology-forecast-challenge-open", "team_id": 13812503, "solution_url": "/competitions/geology-forecast-challenge-open/writeups/abdulkareem-omer-public-8th-private-13th-solution" }