【LightGBM】shapライブラリ無しでSHAP値を出す（pred_contribの使い方）

import numpy as np
import lightgbm as lgb

rng = np.random.RandomState(42)
n = 3000

area  = rng.uniform(15, 80, n)     # 広さ(m2)
age   = rng.uniform(0, 40, n)      # 築年数
dist  = rng.uniform(1, 30, n)      # 駅からの距離(分)
floor = rng.randint(1, 15, n)      # 階数

# 広い/新しい/駅近/高層 ほど家賃が高い
y = 3 + 0.12*area - 0.06*age - 0.15*dist + 0.1*floor + rng.randn(n)*0.5

X = np.column_stack([area, age, dist, floor])
names = ["area", "age", "dist", "floor"]

dtr = lgb.Dataset(X, y, feature_name=names)
model = lgb.train(dict(objective="regression", num_leaves=31, learning_rate=0.05,
                       min_child_samples=20, verbose=-1),
                  dtr, num_boost_round=300)

contrib = model.predict(X, pred_contrib=True)
print(contrib.shape)   # => (3000, 5)

pred = model.predict(X)
row_sum = contrib.sum(axis=1)

print("最大絶対差 |行合計 - predict|:", np.max(np.abs(row_sum - pred)).round(6))
print("ベース値:", round(contrib[0, -1], 4), "/ y の平均:", round(y.mean(), 4))

i = 1
print("特徴量値:", dict(zip(names, np.round(X[i], 1))))
for nm, c in zip(names, contrib[i, :-1]):
    print(f"  {nm:6s} 寄与 {c:+.3f}")
print(f"  base   {contrib[i, -1]:+.3f}")
print(f"  合計 = {contrib[i].sum():.3f}  (predict={pred[i]:.3f})")

import matplotlib.pyplot as plt

base = contrib[i, -1]
contribs = contrib[i, :-1]
running = base

fig, ax = plt.subplots(figsize=(9, 4.8))
ax.bar(0, base, color="#888888")
for k, c in enumerate(contribs, start=1):
    color = "#E45756" if c >= 0 else "#4C78A8"   # 赤=押し上げ, 青=押し下げ
    ax.bar(k, c, bottom=running, color=color)
    running += c
ax.bar(len(names) + 1, pred[i], color="#333333")
ax.set_xticks(range(len(names) + 2))
ax.set_xticklabels(["base"] + names + ["prediction"])
ax.set_ylabel("rent (predicted)")
ax.axhline(base, color="#bbbbbb", ls="--", lw=0.8)
plt.tight_layout()
plt.show()

mean_abs = np.abs(contrib[:, :-1]).mean(axis=0)   # ベース値の列は除く

for nm, v in sorted(zip(names, mean_abs), key=lambda t: -t[1]):
    print(f"{nm:6s} {v:.3f}")

import matplotlib.pyplot as plt

order = np.argsort(mean_abs)
plt.barh([names[i] for i in order], [mean_abs[i] for i in order])
plt.xlabel("mean |SHAP value|")
plt.show()

import matplotlib.pyplot as plt

contribs = model.predict(X, pred_contrib=True)[:, :-1]  # ベース値の列は除く
order = np.argsort(np.abs(contribs).mean(0))            # 重要な特徴量ほど上に

SP, CAP = 0.017, 0.45   # 点の縦間隔（一定）と、行同士がぶつからないための上限

fig, ax = plt.subplots(figsize=(8, 4.6))
for row, fi in enumerate(order):
    sv, fv = contribs[:, fi], X[:, fi]
    color = (fv - fv.min()) / (fv.max() - fv.min() + 1e-9)  # 特徴量値を0-1に
    # SHAP値をビンに分け、同じビンの点を一定間隔で積む → 密度が高いほど縦に膨らむ
    jit = np.zeros(len(sv))
    idx = np.digitize(sv, np.linspace(sv.min(), sv.max(), 120))
    for b in np.unique(idx):
        m = np.where(idx == b)[0]
        off = (np.arange(len(m)) - (len(m)-1)/2) * SP
        jit[m] = np.clip(off, -CAP, CAP)
    ax.scatter(sv, row + jit, c=color, cmap="coolwarm", s=5, alpha=0.7, linewidths=0)

ax.axvline(0, color="#999999", lw=0.8)
ax.set_yticks(range(len(order))); ax.set_yticklabels([names[i] for i in order])
ax.set_xlabel("SHAP value (impact on prediction)")
sm = plt.cm.ScalarMappable(cmap="coolwarm"); sm.set_array([])
cb = fig.colorbar(sm, ax=ax, ticks=[0, 1]); cb.set_ticklabels(["low", "high"])
cb.set_label("feature value")
plt.tight_layout(); plt.show()