P3-4.2 一旦样本单位摇摆，哪些东西会一起摇摆¶

# 这个例子检查样本单位摇摆时，特征、标签和基准是否也一起变化。
import pandas as pd

raw = pd.DataFrame(
    [
        {"event_id": "A", "second": 0, "flow": 0.5, "review_needed": 1},
        {"event_id": "A", "second": 1, "flow": 1.8, "review_needed": 1},
        {"event_id": "A", "second": 2, "flow": 1.1, "review_needed": 1},
        {"event_id": "B", "second": 0, "flow": 0.4, "review_needed": 0},
        {"event_id": "B", "second": 1, "flow": 1.1, "review_needed": 0},
        {"event_id": "B", "second": 2, "flow": 1.0, "review_needed": 0},
        {"event_id": "C", "second": 0, "flow": 0.6, "review_needed": 1},
        {"event_id": "C", "second": 1, "flow": 1.9, "review_needed": 1},
        {"event_id": "C", "second": 2, "flow": 1.3, "review_needed": 1},
    ]
)

per_row = raw[["event_id", "second", "flow", "review_needed"]]
per_event = (
    raw.groupby("event_id", as_index=False)
    .agg(
        flow_mean=("flow", "mean"),
        flow_max=("flow", "max"),
        late_drop=("flow", lambda s: s.iloc[-2] - s.iloc[-1]),
        review_needed=("review_needed", "max"),
    )
)

row_split = raw.assign(split=lambda df: df["second"].map({0: "train", 1: "train", 2: "test"}))
event_split = per_event.assign(split=lambda df: df["event_id"].map({"A": "train", "B": "train", "C": "test"}))

unit_summary = pd.DataFrame(
    [
        {
            "unit": "row",
            "sample_count": len(per_row),
            "feature_example": "flow at one second",
            "label_rows": per_row["review_needed"].sum(),
            "train_events": ",".join(sorted(row_split.loc[row_split["split"] == "train", "event_id"].unique())),
            "test_events": ",".join(sorted(row_split.loc[row_split["split"] == "test", "event_id"].unique())),
        },
        {
            "unit": "event",
            "sample_count": len(per_event),
            "feature_example": "flow_mean / flow_max / late_drop",
            "label_rows": per_event["review_needed"].sum(),
            "train_events": ",".join(sorted(event_split.loc[event_split["split"] == "train", "event_id"].unique())),
            "test_events": ",".join(sorted(event_split.loc[event_split["split"] == "test", "event_id"].unique())),
        },
    ]
)

print("1) sample counts change with the chosen unit")
print("row-level samples:", len(per_row))
print("event-level samples:", len(per_event))
print()
print("2) row-level labels are repeated because the label belongs to the whole event")
print(per_row.groupby("event_id", as_index=False)["review_needed"].sum())
print()
print("3) event-level features and labels line up on the same unit")
print(per_event)
print()
print("4) split stability differs by unit")
print(unit_summary)

1) sample counts change with the chosen unit
row-level samples: 9
event-level samples: 3

2) row-level labels are repeated because the label belongs to the whole event
  event_id  review_needed
0        A              3
1        B              0
2        C              3

3) event-level features and labels line up on the same unit
  event_id  flow_mean  flow_max  late_drop  review_needed
0        A   1.133333       1.8        0.7              1
1        B   0.833333       1.1        0.1              0
2        C   1.266667       1.9        0.6              1

4) split stability differs by unit
    unit  sample_count                    feature_example  label_rows train_events test_events
0    row             9                 flow at one second           6        A,B,C       A,B,C
1  event             3  flow_mean / flow_max / late_drop           2          A,B           C

# 这个例子用来确认同一份原始日志的分割单位会如何改变评估分数。
import pandas as pd
from sklearn.metrics import accuracy_score
from sklearn.tree import DecisionTreeClassifier

raw_rows = []
labels = {"A": 1, "B": 0, "C": 1, "D": 0, "E": 1, "F": 0, "G": 0, "H": 1}
base_flow = {"A": 10, "B": 30, "C": 50, "D": 70, "E": 90, "F": 110, "G": 130, "H": 150}

for event_id, label in labels.items():
    for second, offset in enumerate([0.0, 0.2, -0.1]):
        raw_rows.append(
            {
                "event_id": event_id,
                "second": second,
                "flow": base_flow[event_id] + offset,
                "review_needed": label,
            }
        )

raw = pd.DataFrame(raw_rows)

# 行级分割：每个 event_id 的部分行会同时进入 train 和 test。
row_train = raw[raw["second"].isin([0, 1])]
row_test = raw[raw["second"].eq(2)]

# 动作级分割：test event_id 会从训练中被完整排除。
event_train = raw[raw["event_id"].isin(["A", "B", "C", "D"])]
event_test = raw[raw["event_id"].isin(["E", "F", "G", "H"])]


def evaluate(train, test):
    model = DecisionTreeClassifier(random_state=0)
    model.fit(train[["flow"]], train["review_needed"])
    predictions = model.predict(test[["flow"]])
    return accuracy_score(test["review_needed"], predictions), [
        (event_id, int(prediction), int(actual))
        for event_id, prediction, actual in zip(test["event_id"], predictions, test["review_needed"])
    ]


row_accuracy, _ = evaluate(row_train, row_test)
event_accuracy, event_predictions = evaluate(event_train, event_test)
leaked_events = sorted(set(row_train["event_id"]) & set(row_test["event_id"]))

print("leaked events in row split:", leaked_events)
print("row split accuracy:", row_accuracy)
print("event split accuracy:", event_accuracy)
print("event split predictions:", event_predictions)

leaked events in row split: ['A', 'B', 'C', 'D', 'E', 'F', 'G', 'H']
row split accuracy: 1.0
event split accuracy: 0.5
event split predictions: [('E', 0, 1), ('E', 0, 1), ('E', 0, 1), ('F', 0, 0), ('F', 0, 0), ('F', 0, 0), ('G', 0, 0), ('G', 0, 0), ('G', 0, 0), ('H', 0, 1), ('H', 0, 1), ('H', 0, 1)]