大事な事を忘れてた

モデルの指数は書いてるけど、どう設定してそれが得られたか書いてないので更に攻めるにしてもどんな設定試してたのか覚えてないorz まあ、イテレーション数5,000なのはモデルのファイル名で分かるけど、それ以外は全く不明ではあるが、それでもまあ、葉255、率は0.01～0.02の間を攻めてたと思うし、分割数は63辺りをやってたかと思うので、今晩も残業後だけど少し試そうかと、イテレーション数5,000、葉271、率0.01、分割数47、葉の最小50、ボーダー500でNDCG@1=0.7817 NDCG@3=0.7861 NDCG@5=0.7993 NDCG@10=0.8533とちょっと詳しく書いてみた。これはログに

=== LightGBM Ranking 学習開始 === (開始時刻: 2026-06-16 20:48:19)
=== データ読み込み開始 ===
=== Train/Test 分割 ===
=== パイプライン構築 ===
=== 学習開始 === 2026-06-16 20:48:19
NumberOfLeaves: 271
NumberOfIterations: 5000
LearningRate: 0.01
MaximumBinCountPerFeature: 47
MinimumExampleCountPerLeaf: 50
CustomGains: 0～500 = 0点, 501～1000はリニアに0～1000点
=== 学習完了 === 2026-06-16 20:59:52(00:11:33.1451069)
=== 評価（自前NDCG計算）開始 === 2026-06-16 20:59:52
NDCG@1 = 0.7817
NDCG@3 = 0.7861
NDCG@5 = 0.7993
NDCG@10 = 0.8533
=== 評価完了 === 2026-06-16 21:02:02 (00:02:09.5755349)
=== モデル保存 ===
保存完了: D:\SaraD\ML\RankCK5000_NDCG7817_7861.zip
=== LightGBM Ranking 学習完了 === (終了時刻: 2026-06-16 21:02:04 所要時間: 00:13:45.4878322)

って感じに表示するようにしてます。

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.29% (79.17%)	26.70% (79.31%)	27.58% (78.40%)	31.67% (87.75%)	36.07% (83.01%)	29.17% (82.79%)	25.88% (77.62%)	58.52% (78.99%)
複勝	59.07% (85.01%)	58.97% (84.99%)	58.57% (84.69%)	67.50% (89.67%)	69.86% (86.62%)	63.54% (84.40%)	56.61% (85.07%)	89.35% (83.06%)
枠連	16.16% (84.99%)	16.68% (96.00%)	15.12% (73.81%)	25.25% (106.06%)	-- (--)	17.68% (78.42%)	14.63% (81.79%)	31.57% (80.40%)
馬連	12.50% (80.53%)	12.91% (85.23%)	11.54% (74.09%)	20.00% (102.50%)	22.83% (62.05%)	14.34% (72.47%)	10.96% (84.88%)	26.02% (74.89%)
ワイド	27.61% (86.56%)	28.77% (88.08%)	25.76% (83.35%)	36.67% (109.08%)	50.23% (83.20%)	32.88% (85.18%)	23.82% (87.32%)	49.55% (81.07%)
馬単	6.77% (68.71%)	6.69% (75.93%)	6.50% (59.17%)	11.67% (98.08%)	12.33% (60.37%)	7.91% (65.82%)	5.89% (70.43%)	26.02% (72.02%)
三連複	6.51% (85.52%)	7.34% (79.24%)	5.47% (91.65%)	9.17% (89.83%)	16.89% (68.36%)	8.78% (71.46%)	4.82% (91.76%)	17.28% (86.01%)
三連単	1.27% (67.95%)	1.54% (110.85%)	1.09% (28.88%)	0.00% (0.00%)	3.20% (39.45%)	1.98% (45.28%)	0.87% (78.08%)	17.28% (79.91%)
総合	60.41% (79.73%)	60.51% (87.27%)	59.72% (71.75%)	68.33% (84.91%)	74.43% (69.01%)	64.40% (73.23%)	57.81% (82.12%)	89.67% (79.36%)

これも微妙だ^^; 葉255に戻して率0.008、分割数63に戻す。NDCG@1=0.7813 NDCG@3=0.7853 NDCG@5=0.7986 NDCG@10=0.8530と全てダウンでダメダメなのは承知なんだが、何故こんな選択してみたかは、実は多分メモリギリギリなのが分かってるからで、案の定、連続で検証しようとするとOOMです。一旦アプリ閉じて再起動すれば検証出来てるけど、それは好ましい事じゃないんだよね。もって軽いモデルで精度を上げなきゃと思うから多分新たな特徴量を検討しなきゃダメかも。

3,455R	1点	芝(1,689R)	ダート(1,646R)	障害(120R)	8頭以下(219R)	9～12頭(809R)	13頭以上(2,427R)	多点
単勝	27.79% (81.50%)	27.59% (82.29%)	27.64% (79.09%)	32.50% (103.42%)	36.99% (85.30%)	29.67% (83.52%)	26.33% (80.48%)	58.03% (76.23%)
複勝	59.02% (84.74%)	58.79% (84.33%)	58.81% (84.80%)	65.00% (89.67%)	71.69% (89.63%)	64.03% (84.68%)	56.20% (84.32%)	89.29% (82.74%)
枠連	16.39% (87.16%)	16.82% (98.02%)	15.57% (76.39%)	23.23% (103.84%)	-- (--)	17.80% (81.09%)	14.87% (83.65%)	30.82% (74.48%)
馬連	12.62% (80.25%)	13.08% (83.86%)	11.73% (75.30%)	18.33% (97.25%)	22.83% (58.13%)	15.08% (82.22%)	10.88% (81.59%)	25.56% (70.09%)
ワイド	27.29% (84.85%)	28.72% (86.32%)	25.58% (83.59%)	30.83% (81.33%)	49.32% (80.73%)	32.14% (82.00%)	23.69% (86.17%)	50.07% (79.02%)
馬単	6.83% (79.21%)	6.93% (83.96%)	6.44% (71.10%)	10.83% (123.50%)	11.42% (48.77%)	8.16% (88.34%)	5.97% (78.91%)	25.56% (66.79%)
三連複	6.05% (70.41%)	7.05% (67.29%)	4.98% (74.17%)	6.67% (62.83%)	19.18% (68.26%)	8.53% (69.33%)	4.04% (70.97%)	17.05% (79.87%)
三連単	1.24% (66.10%)	1.42% (98.03%)	1.15% (38.15%)	0.00% (0.00%)	2.74% (33.56%)	1.85% (35.97%)	0.91% (79.08%)	17.05% (74.33%)
総合	60.23% (79.17%)	60.04% (85.25%)	60.02% (72.80%)	65.83% (82.26%)	75.80% (66.34%)	64.52% (75.90%)	57.40% (80.64%)	89.61% (74.53%)

まあ、落ちてはいるけど覚悟の上で、どっちのモデルもリリースには不向きなのは分かった。