# 後段モデル比較結果

- 入力: `local_batch_api/output_full/training_dataset.csv`
- 件数: 519
- 100点スケール外で除外: 1
- assessable: 346 / 519
- 判定不可含む行: 173 / 519
- split: train/val/test = 70/10/20, stratified by `human_total_score`, seed=42

## 出力ファイル

- `features.csv`: モデル投入用特徴量
- `splits.csv`: 固定分割
- `model_comparison.csv`: 6モデル比較結果
- `predictions.csv`: 各動画ごとの予測と誤差
- `feature_importance.csv`: 重要特徴量

## 上位結果(test MAE順)

| target | feature_set | model | val_n | val_rho | val_mae | val_rmse | val_grade_match | val_anomaly_recall | val_anomaly_n | test_n | test_rho | test_mae | test_rmse | test_grade_match | test_anomaly_recall | test_anomaly_n |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| human_soup_score | visual_only | 線形回帰 | 52 | 0.4742 | 8.687 | 11.28 | 0.6731 | 0 | 1 | 104 | 0.5188 | 8.815 | 11.65 | 0.7308 | 0 | 4 |
| human_soup_score | visual_plus_brix_slot | 線形回帰 | 52 | 0.4648 | 8.845 | 11.34 | 0.6731 | 0 | 1 | 104 | 0.5272 | 8.834 | 11.6 | 0.7596 | 0 | 4 |
| human_soup_score | visual_plus_brix | 線形回帰 | 52 | 0.4722 | 8.846 | 11.34 | 0.6731 | 0 | 1 | 104 | 0.5252 | 8.834 | 11.6 | 0.7596 | 0 | 4 |
| human_soup_score | visual_only | Ridge回帰 | 52 | 0.4018 | 8.793 | 11.36 | 0.7115 | 0 | 1 | 104 | 0.511 | 8.931 | 12.03 | 0.7308 | 0 | 4 |
| human_soup_score | visual_plus_brix_slot | Ridge回帰 | 52 | 0.4312 | 8.975 | 11.5 | 0.7115 | 0 | 1 | 104 | 0.5399 | 8.942 | 12 | 0.7692 | 0 | 4 |
| human_total_score | visual_plus_brix_slot | Random Forest | 52 | 0.4449 | 10.04 | 11.91 | 0.6731 | 0 | 1 | 104 | 0.6196 | 8.715 | 11.33 | 0.8077 | 0 | 2 |
| human_total_score | visual_plus_brix_slot | LightGBM | 52 | 0.4716 | 9.654 | 11.38 | 0.7115 | 0 | 1 | 104 | 0.634 | 8.72 | 11.46 | 0.7885 | 0 | 2 |
| human_total_score | visual_plus_brix | 線形回帰 | 52 | 0.4991 | 9.425 | 11.4 | 0.75 | 0 | 1 | 104 | 0.6224 | 8.75 | 11.07 | 0.8077 | 0 | 2 |
| human_total_score | visual_plus_brix_slot | 線形回帰 | 52 | 0.4921 | 9.427 | 11.4 | 0.75 | 0 | 1 | 104 | 0.6154 | 8.751 | 11.07 | 0.8077 | 0 | 2 |
| human_total_score | visual_plus_brix_slot | Ridge回帰 | 52 | 0.532 | 9.307 | 11.32 | 0.7308 | 0 | 1 | 104 | 0.6333 | 8.824 | 11.36 | 0.8077 | 0 | 2 |

## 注意

- 人間コメントは特徴量に入れていません。
- 固定配点は学習なしの基準値です。
- `visual_density` は既存出力を使用し、評価方針自体は変更していません。