In the final stage, when I was running out of ideas and trying to adjust parameters, I discovered an improvement method whose specific mechanism I wasn't clear about: attempting to ”polarize“ the differences. Actually, I didn't think it could improve the scores - I just regarded it as another novel but low-precision way of \"removing noise\", and didn't expect it to bring any benefits.
\nWho can tell me the reason why this method brings profits? I'm still feeling confused about it.
", "messages": [ { "id": "3400295", "postDate": "02/01/2026 10:53:19", "content": "In the final stage, when I was running out of ideas and trying to adjust parameters, I discovered an improvement method whose specific mechanism I wasn't clear about: attempting to ”polarize“ the differences. Actually, I didn't think it could improve the scores - I just regarded it as another novel but low-precision way of \"removing noise\", and didn't expect it to bring any benefits.
\nWho can tell me the reason why this method brings profits? I'm still feeling confused about it.
", "rawMarkdown": "In the final stage, when I was running out of ideas and trying to adjust parameters, I discovered an improvement method whose specific mechanism I wasn't clear about: attempting to ”polarize“ the differences. Actually, I didn't think it could improve the scores - I just regarded it as another novel but low-precision way of \"removing noise\", and didn't expect it to bring any benefits.\n\nWho can tell me the reason why this method brings profits? I'm still feeling confused about it.", "votes": null }, { "id": "3400359", "postDate": "02/01/2026 12:59:00", "content": "In principle, if a strictly monotonic transformation is applied equally to all scores, it should not affect the evaluation metric beyond a shift in the optimal threshold, since the relative ranking is preserved. Any observed improvement would therefore come from how the transformation interacts with thresholding, score distributions, or the merging of calibrated predictors, rather than from adding new information.
", "rawMarkdown": "In principle, if a strictly monotonic transformation is applied equally to all scores, it should not affect the evaluation metric beyond a shift in the optimal threshold, since the relative ranking is preserved. Any observed improvement would therefore come from how the transformation interacts with thresholding, score distributions, or the merging of calibrated predictors, rather than from adding new information.", "votes": null }, { "id": "3400418", "postDate": "02/01/2026 14:15:14", "content": "Let me put it this way. I just slightly increased the larger scores and decreased the smaller ones, and set some parameters, and then I got a 0.002 to 0.004 improvement. I don't understand.
", "rawMarkdown": "Let me put it this way. I just slightly increased the larger scores and decreased the smaller ones, and set some parameters, and then I got a 0.002 to 0.004 improvement. I don't understand.", "votes": null }, { "id": "3400465", "postDate": "02/01/2026 15:40:29", "content": "I got your point. I'm not totally sure why, but it's probably related to why people submit scores larger than 1.
", "rawMarkdown": "I got your point. I'm not totally sure why, but it's probably related to why people submit scores larger than 1.", "votes": null } ], "comments": [ { "id": 3400359, "author_name": "xaxipirulazo", "author_url": "", "post_date": "02/01/2026 12:59:00", "content": "In principle, if a strictly monotonic transformation is applied equally to all scores, it should not affect the evaluation metric beyond a shift in the optimal threshold, since the relative ranking is preserved. Any observed improvement would therefore come from how the transformation interacts with thresholding, score distributions, or the merging of calibrated predictors, rather than from adding new information.
", "votes": null, "replies": [ { "id": 3400418, "author_name": "zejunfool", "author_url": "", "post_date": "02/01/2026 14:15:14", "content": "Let me put it this way. I just slightly increased the larger scores and decreased the smaller ones, and set some parameters, and then I got a 0.002 to 0.004 improvement. I don't understand.
", "votes": null, "replies": [ { "id": 3400465, "author_name": "xaxipirulazo", "author_url": "", "post_date": "02/01/2026 15:40:29", "content": "I got your point. I'm not totally sure why, but it's probably related to why people submit scores larger than 1.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3400295": "In the final stage, when I was running out of ideas and trying to adjust parameters, I discovered an improvement method whose specific mechanism I wasn't clear about: attempting to ”polarize“ the differences. Actually, I didn't think it could improve the scores - I just regarded it as another novel but low-precision way of \"removing noise\", and didn't expect it to bring any benefits.\n\nWho can tell me the reason why this method brings profits? I'm still feeling confused about it.", "3400359": "In principle, if a strictly monotonic transformation is applied equally to all scores, it should not affect the evaluation metric beyond a shift in the optimal threshold, since the relative ranking is preserved. Any observed improvement would therefore come from how the transformation interacts with thresholding, score distributions, or the merging of calibrated predictors, rather than from adding new information.", "3400418": "Let me put it this way. I just slightly increased the larger scores and decreased the smaller ones, and set some parameters, and then I got a 0.002 to 0.004 improvement. I don't understand.", "3400465": "I got your point. I'm not totally sure why, but it's probably related to why people submit scores larger than 1." }, "source": "meta" }