Thanks for organizing this interesting yet challenging competition with real-world data targeting life-threatening cancer.
\nI learned a lot by participating, and though I couldn't reach top ranks, I'm still happy with my results. I wanted to note down my takeaways, so I can refer back to them when I join a new competition. Then I thought, why not share them here so others may find them interesting and useful.
\nIt is crucial to understand the data in detail. Gaining a little domain knowledge (in this case, medical knowledge and how doctors actually annotate these pictures) could be helpful.
\nPlan your experiments, what you want to explore or learn with consideration for timelines. Organize and schedule your ideas systematically to ensure you explore all possibilities. I personally put lots of time on testing if straight forward classification works or not and didn't have time to try MIL.
\nTry to not miss any discussion, insights from the community can be invaluable in refining your approach and improving your models.
\nAvoid using codes that you don't fully understand. Tiny but vital details can be ignored which cause lots of wasting time in wrong direction.
\nKeep detailed records of your experiments. Even though there are tons of tools for it, just writing them somewhere can give you a big-picture view of what you are doing, enabling better comparisons between your experiments.
\nI realized that emotional attachment to the leaderboard rankings is counterproductive. The focus should be on learning and understanding what works and what doesn't, and this process eventually improves your LB score.
\nIt really works!
", "messages": [ { "id": 2587630, "postDate": "2024-01-04T22:21:54.673Z", "content": "Thanks for organizing this interesting yet challenging competition with real-world data targeting life-threatening cancer.
\nI learned a lot by participating, and though I couldn't reach top ranks, I'm still happy with my results. I wanted to note down my takeaways, so I can refer back to them when I join a new competition. Then I thought, why not share them here so others may find them interesting and useful.
\nIt is crucial to understand the data in detail. Gaining a little domain knowledge (in this case, medical knowledge and how doctors actually annotate these pictures) could be helpful.
\nPlan your experiments, what you want to explore or learn with consideration for timelines. Organize and schedule your ideas systematically to ensure you explore all possibilities. I personally put lots of time on testing if straight forward classification works or not and didn't have time to try MIL.
\nTry to not miss any discussion, insights from the community can be invaluable in refining your approach and improving your models.
\nAvoid using codes that you don't fully understand. Tiny but vital details can be ignored which cause lots of wasting time in wrong direction.
\nKeep detailed records of your experiments. Even though there are tons of tools for it, just writing them somewhere can give you a big-picture view of what you are doing, enabling better comparisons between your experiments.
\nI realized that emotional attachment to the leaderboard rankings is counterproductive. The focus should be on learning and understanding what works and what doesn't, and this process eventually improves your LB score.
\nIt really works!
", "rawMarkdown": "Thanks for organizing this interesting yet challenging competition with real-world data targeting life-threatening cancer.\n\nI learned a lot by participating, and though I couldn't reach top ranks, I'm still happy with my results. I wanted to note down my takeaways, so I can refer back to them when I join a new competition. Then I thought, why not share them here so others may find them interesting and useful.\n\n### Understand Data \nIt is crucial to understand the data in detail. Gaining a little domain knowledge (in this case, medical knowledge and how doctors actually annotate these pictures) could be helpful.\n\n### Timeline \nPlan your experiments, what you want to explore or learn with consideration for timelines. Organize and schedule your ideas systematically to ensure you explore all possibilities. I personally put lots of time on testing if straight forward classification works or not and didn't have time to try MIL.\n\n### Discussions\nTry to not miss any discussion, insights from the community can be invaluable in refining your approach and improving your models.\n\n### Understanding Code Notebooks\nAvoid using codes that you don't fully understand. Tiny but vital details can be ignored which cause lots of wasting time in wrong direction.\n\n### Experiments Tracking\nKeep detailed records of your experiments. Even though there are tons of tools for it, just writing them somewhere can give you a big-picture view of what you are doing, enabling better comparisons between your experiments.\n\n### Leaderboard \nI realized that emotional attachment to the leaderboard rankings is counterproductive. The focus should be on learning and understanding what works and what doesn't, and this process eventually improves your LB score.\n\n### Model Ensemble \nIt really works!", "votes": 5 } ], "comments": [], "raw_markdown_by_id": { "2587630": "Thanks for organizing this interesting yet challenging competition with real-world data targeting life-threatening cancer.\n\nI learned a lot by participating, and though I couldn't reach top ranks, I'm still happy with my results. I wanted to note down my takeaways, so I can refer back to them when I join a new competition. Then I thought, why not share them here so others may find them interesting and useful.\n\n### Understand Data \nIt is crucial to understand the data in detail. Gaining a little domain knowledge (in this case, medical knowledge and how doctors actually annotate these pictures) could be helpful.\n\n### Timeline \nPlan your experiments, what you want to explore or learn with consideration for timelines. Organize and schedule your ideas systematically to ensure you explore all possibilities. I personally put lots of time on testing if straight forward classification works or not and didn't have time to try MIL.\n\n### Discussions\nTry to not miss any discussion, insights from the community can be invaluable in refining your approach and improving your models.\n\n### Understanding Code Notebooks\nAvoid using codes that you don't fully understand. Tiny but vital details can be ignored which cause lots of wasting time in wrong direction.\n\n### Experiments Tracking\nKeep detailed records of your experiments. Even though there are tons of tools for it, just writing them somewhere can give you a big-picture view of what you are doing, enabling better comparisons between your experiments.\n\n### Leaderboard \nI realized that emotional attachment to the leaderboard rankings is counterproductive. The focus should be on learning and understanding what works and what doesn't, and this process eventually improves your LB score.\n\n### Model Ensemble \nIt really works!" } }