https://horikitasaku.github.io/en/tags/lightgbm/Recent content in LightGBM on YuXuan Wu/HorikitaSakuHugo 0.125.0en&copy; 2023 <a href="">Horikita Saku</a>Mon, 11 Sep 2023 00:00:00 +0000https://horikitasaku.github.io/en/posts/fda-cosine-similarity/Mon, 11 Sep 2023 00:00:00 +0000https://horikitasaku.github.io/en/posts/fda-cosine-similarity/Foreword Nothing fancy here — I just felt like writing something. In a Nishika competition, yama brought up the Negative Cosine Similarity Loss and wired it up as a custom loss for lgbm. → yama&rsquo;s notebook This article is also posted on Qiita. The finite difference trick and how it gets applied there struck me as genuinely interesting, so I wanted to make it the subject of my first blog post. I won&rsquo;t get into how well the loss actually performs in that competition.