ランキング学習のメモ

先日のリリースに添付したランキング学習のモデルはRankCK3000_loop1_NDCG7679_7741.zipで

2,610R	1点	芝(1,278R)	ダート(1,243R)	障害(89R)	8頭以下(157R)	9～12頭(665R)	13頭以上(1,788R)	多点
単勝	24.75% (75.89%)	25.04% (81.96%)	24.70% (71.30%)	21.35% (52.81%)	33.76% (74.78%)	27.07% (71.82%)	23.10% (77.51%)	54.67% (75.35%)
複勝	55.52% (80.09%)	53.36% (78.44%)	57.52% (82.08%)	58.43% (76.07%)	64.33% (92.36%)	60.45% (78.06%)	52.91% (79.77%)	89.66% (81.16%)
枠連	13.76% (74.67%)	12.13% (75.46%)	14.90% (72.29%)	18.99% (99.24%)	-- (--)	13.98% (60.81%)	12.86% (75.44%)	30.49% (76.91%)
馬連	9.96% (65.54%)	8.69% (66.33%)	11.10% (64.22%)	12.36% (72.70%)	20.38% (87.83%)	11.43% (63.91%)	8.50% (64.19%)	23.60% (71.87%)
ワイド	24.18% (73.08%)	22.69% (71.83%)	25.91% (75.45%)	21.35% (58.09%)	40.13% (78.85%)	27.22% (70.56%)	21.64% (73.52%)	46.17% (76.18%)
馬単	5.48% (56.70%)	4.93% (58.16%)	6.11% (56.11%)	4.49% (43.93%)	12.10% (81.21%)	6.77% (66.80%)	4.42% (50.79%)	23.60% (70.33%)
三連複	6.59% (71.77%)	6.34% (68.12%)	7.00% (78.69%)	4.49% (27.64%)	13.38% (90.45%)	7.37% (61.35%)	5.70% (74.01%)	16.93% (77.84%)
三連単	1.46% (90.74%)	1.02% (49.79%)	1.93% (137.14%)	1.12% (30.67%)	1.91% (106.43%)	1.65% (58.74%)	1.34% (101.26%)	16.93% (74.78%)
総合	57.20% (73.55%)	55.56% (68.63%)	58.81% (79.71%)	58.43% (57.05%)	67.52% (87.42%)	60.90% (66.51%)	54.92% (74.56%)	89.96% (74.93%)

この際に使ったパラメーターが分からなくなった(笑) イテレーション数3,000なのは分かるが、率は多分0.01だったのかな^^; モデルサイズが21MBなのでそんな気がするけど。

パラメータ / 指標	条件 1	条件 2	条件 3	条件 4	条件 5	条件 6	条件 7
NumberOfLeaves	79	95	95	111	127	255	351
NumberOfIterations	3,000	3,000	3,000	3,000	3,000	3,000	3,000
LearningRate	0.010	0.010	0.005	0.010	0.010	0.010	0.010
MaximumBin	127	127	127	127	127	127	127
MinimumExample	50	50	50	50	50	50	50
SubsampleFraction	0.800	0.800	0.800	0.800	0.800	0.800	0.800
FeatureFraction	0.75	0.75	0.75	0.75	0.75	0.75	0.75
ボーダー	300	300	300	300	300	300	300
乗数	0.95	0.95	0.95	0.95	0.95	0.95	0.95
NDCG@1	0.7667	0.7660	0.7635	0.7673	0.7669	0.7668	0.7665
NDCG@3	0.7725	0.7723	0.7701	0.7730	0.7731	0.7731	0.7736
NDCG@5	0.7889	0.7891	0.7867	0.7893	0.7895	0.7893	0.7895
NDCG@10	0.8498	0.8499	0.8482	0.8502	0.8504	0.8505	0.8507

なんか以前、葉を増やして出したモデルっぽいな。最近の走破タイムでは葉の数を絞ってたので忘れてるけど300以上も試してた記憶もあるからなぁ。

NumberOfLeaves: 335
NumberOfIterations: 3000
LearningRate: 0.01
MaximumBinCountPerFeature: 127
MinimumExampleCountPerLeaf: 50
SubsampleFraction: 0.8
FeatureFraction: 0.75
CustomGains: 0～300 = 0点, 301～1000は0.95乗に0～1000点

これ10回試したら最後に

[Loop 10] NDCG@1 = 0.7683
[Loop 10] NDCG@3 = 0.7740
[Loop 10] NDCG@5 = 0.7902
[Loop 10] NDCG@10 = 0.8509

が出た。

2,610R	1点	芝(1,278R)	ダート(1,243R)	障害(89R)	8頭以下(157R)	9～12頭(665R)	13頭以上(1,788R)	多点
単勝	24.37% (74.99%)	23.87% (74.20%)	25.02% (77.40%)	22.47% (52.58%)	34.39% (74.39%)	27.82% (71.64%)	22.20% (76.29%)	54.90% (75.18%)
複勝	54.90% (78.89%)	52.66% (76.56%)	57.12% (81.78%)	56.18% (72.02%)	64.33% (85.92%)	59.70% (76.42%)	52.29% (79.19%)	89.77% (81.65%)
枠連	13.67% (76.25%)	12.13% (81.24%)	14.73% (69.86%)	18.99% (103.16%)	-- (--)	13.38% (58.06%)	12.98% (78.54%)	30.28% (78.02%)
馬連	10.04% (64.96%)	9.00% (64.98%)	10.94% (64.22%)	12.36% (75.06%)	21.02% (80.89%)	10.83% (61.10%)	8.78% (64.99%)	23.52% (74.94%)
ワイド	24.44% (73.08%)	23.47% (72.16%)	25.66% (75.07%)	21.35% (58.43%)	40.76% (77.45%)	27.37% (72.23%)	21.92% (73.01%)	46.51% (79.31%)
馬単	5.56% (59.32%)	4.69% (49.66%)	6.36% (69.24%)	6.74% (59.44%)	12.10% (68.79%)	6.92% (62.26%)	4.47% (57.40%)	23.52% (72.10%)
三連複	7.05% (77.72%)	7.12% (78.04%)	7.16% (80.98%)	4.49% (27.64%)	15.29% (116.94%)	7.52% (64.41%)	6.15% (79.23%)	16.36% (76.34%)
三連単	1.61% (93.76%)	1.25% (57.75%)	2.01% (135.30%)	1.12% (30.67%)	3.82% (181.27%)	1.95% (49.79%)	1.29% (102.43%)	16.36% (71.54%)
総合	56.67% (74.85%)	54.85% (69.10%)	58.49% (81.81%)	57.30% (59.26%)	68.15% (97.95%)	60.30% (64.49%)	54.31% (76.38%)	90.04% (73.89%)

やはり、この辺りなんだなぁ^^;