P3-6.1 应该用什么特征把可比较的结构留下来¶

# 这个例子从区段汇总值中生成并选择符合比较目的的特征。
from statistics import mean, stdev

feature_focus = "change"

segment_summary = [
    {"event_id": "A", "early_flow_mean": 1.8, "mid_flow_mean": 2.2, "late_flow_mean": 2.6},
    {"event_id": "B", "early_flow_mean": 2.1, "mid_flow_mean": 2.2, "late_flow_mean": 2.3},
]

feature_table = []
for row in segment_summary:
    segment_values = [row["early_flow_mean"], row["mid_flow_mean"], row["late_flow_mean"]]
    late_minus_early = row["late_flow_mean"] - row["early_flow_mean"]
    feature_table.append(
        {
            **row,
            "overall_mean": mean(segment_values),
            "late_minus_early": late_minus_early,
            "early_to_late_slope": late_minus_early / 2,
            "segment_variability": stdev(segment_values),
        }
    )

focus_map = {
    "level": ["overall_mean"],
    "change": ["late_minus_early", "early_to_late_slope"],
    "stability": ["segment_variability"],
}
focus_columns = focus_map[feature_focus]

print("1) segment means before feature design")
print("  event_id  early_flow_mean  mid_flow_mean  late_flow_mean")
for index, row in enumerate(segment_summary):
    print(
        f"{index}        {row['event_id']}              {row['early_flow_mean']:.1f}"
        f"            {row['mid_flow_mean']:.1f}             {row['late_flow_mean']:.1f}"
    )
print()
print("2) designed features for comparison")
print(
    "  event_id  overall_mean  late_minus_early  early_to_late_slope"
    "  segment_variability"
)
for index, row in enumerate(feature_table):
    print(
        f"{index}        {row['event_id']}           {row['overall_mean']:.1f}"
        f"               {row['late_minus_early']:.1f}"
        f"                  {row['early_to_late_slope']:.1f}"
        f"                  {row['segment_variability']:.1f}"
    )
print()
print(f"3) selected features when feature_focus = {feature_focus}")
if focus_columns == ["overall_mean"]:
    print("  event_id  overall_mean")
    for index, row in enumerate(feature_table):
        print(f"{index}        {row['event_id']}           {row['overall_mean']:.1f}")
elif focus_columns == ["late_minus_early", "early_to_late_slope"]:
    print("  event_id  late_minus_early  early_to_late_slope")
    for index, row in enumerate(feature_table):
        print(
            f"{index}        {row['event_id']}               {row['late_minus_early']:.1f}"
            f"                  {row['early_to_late_slope']:.1f}"
        )
else:
    print("  event_id  segment_variability")
    for index, row in enumerate(feature_table):
        print(f"{index}        {row['event_id']}                  {row['segment_variability']:.1f}")
print()
print("4) feature_focus comparison")
for focus_name, columns in focus_map.items():
    print(f"- {focus_name}: {columns}")

1) segment means before feature design
  event_id  early_flow_mean  mid_flow_mean  late_flow_mean
0        A              1.8            2.2             2.6
1        B              2.1            2.2             2.3

2) designed features for comparison
  event_id  overall_mean  late_minus_early  early_to_late_slope  segment_variability
0        A           2.2               0.8                  0.4                  0.4
1        B           2.2               0.2                  0.1                  0.1

3) selected features when feature_focus = change
  event_id  late_minus_early  early_to_late_slope
0        A               0.8                  0.4
1        B               0.2                  0.1

4) feature_focus comparison
- level: ['overall_mean']
- change: ['late_minus_early', 'early_to_late_slope']
- stability: ['segment_variability']

# 这个例子比较只保留平均值的模型，以及加入变化/变动性特征的模型预测差异。
import pandas as pd
from sklearn.metrics import accuracy_score
from sklearn.tree import DecisionTreeClassifier

events = pd.DataFrame(
    [
        {"event_id": "A", "early": 1.8, "mid": 2.2, "late": 2.6, "review_needed": 1},
        {"event_id": "B", "early": 2.1, "mid": 2.2, "late": 2.3, "review_needed": 0},
        {"event_id": "C", "early": 2.5, "mid": 2.2, "late": 1.9, "review_needed": 1},
        {"event_id": "D", "early": 2.0, "mid": 2.2, "late": 2.4, "review_needed": 0},
        {"event_id": "E", "early": 1.7, "mid": 2.2, "late": 2.7, "review_needed": 1},
        {"event_id": "F", "early": 2.2, "mid": 2.2, "late": 2.2, "review_needed": 0},
        {"event_id": "G", "early": 2.6, "mid": 2.2, "late": 1.8, "review_needed": 1},
        {"event_id": "H", "early": 2.0, "mid": 2.1, "late": 2.3, "review_needed": 0},
    ]
)

segment_values = events[["early", "mid", "late"]]
events["overall_mean"] = segment_values.mean(axis=1)
events["late_minus_early"] = events["late"] - events["early"]
events["segment_variability"] = segment_values.std(axis=1)

train = events[events["event_id"].isin(["A", "B", "C", "D", "E", "F"])]
test = events[events["event_id"].isin(["G", "H"])]
feature_sets = {
    "mean_only": ["overall_mean"],
    "structure_features": ["overall_mean", "late_minus_early", "segment_variability"],
}

for name, columns in feature_sets.items():
    model = DecisionTreeClassifier(random_state=0, max_depth=2)
    model.fit(train[columns], train["review_needed"])
    predicted = model.predict(test[columns])
    comparison = [
        (event_id, int(prediction), int(actual))
        for event_id, prediction, actual in zip(test["event_id"], predicted, test["review_needed"])
    ]
    print(name, "accuracy:", accuracy_score(test["review_needed"], predicted))
    print(name, "predictions:", comparison)

mean_only accuracy: 0.5
mean_only predictions: [('G', 0, 1), ('H', 0, 0)]
structure_features accuracy: 1.0
structure_features predictions: [('G', 1, 1), ('H', 0, 0)]