KNN 的超参数

Python 中的异常检测

Bekhruz (Bex) Tuychiev

Kaggle Master, Data Science Content Creator

修改 evaluate_outlier_classifier

def evaluate_outlier_classifier(model, data, threshold=.75):
    model.fit(data)

    probs = model.predict_proba(data)
    inliers = data[probs[:, 1] <= threshold]

    return inliers

修改 evaluate_regressor

def evaluate_regressor(inliers):
    X, y = inliers.drop("weightkg", axis=1), inliers[['weightkg']]
    X_train, X_test, y_train, y_test = train_test_split(
      X, y, random_state=10, train_size=0.8)

    lr = LinearRegression()
    lr.fit(X_train, y_train)

    preds = lr.predict(X_test)
    rmse = root_mean_squared_error(y_test, preds)

    return round(rmse, 3)

调参：邻居数量

n_neighbors = [5, 10, 15, 20]
scores = dict()


for k in n_neighbors:
    # 初始化 KNN
    knn = KNN(n_neighbors=k)

    # 用 KNN 获取内点
    inliers = evaluate_outlier_classifier(knn, males_transformed, .55)

    # 计算并存储 RMSE
    scores[k] = evaluate_regressor(inliers)

结果检查

print(scores)

{5: 19.463, 10: 17.965, 15: 18.817, 20: 20.597}

距离度量

支持 40 多种度量

knn = KNN(metric='euclidean')

一个包含 8 个样本的数据集示意图，A 为离群点，并标出 A 与其 4 个最近邻的箭头。

曼哈顿距离

A = np.array([9, 1, 6, ...])
B = np.array([25, 44, 85, ...])


diffs = np.abs(B - A)


manhattan_dist_AB = np.sum(diffs)
manhattan_dist_AB

用于计算曼哈顿距离的公式及其可视化示意。

曼哈顿距离

适用于高维数据
返回值通常大于欧氏距离
更适合类别型特征

用于计算曼哈顿距离的公式及其可视化示意。

闵可夫斯基距离

# 欧氏距离
knn = KNN(metric="minkowski", p=2)

# 曼哈顿距离
knn = KNN(metric="minkowski", p=1)

用于计算闵可夫斯基距离的公式及其可视化示意。

距离聚合方式

knn_largest = KNN(n_neighbors=10, method="largest")


knn_mean = KNN(n_neighbors=10, method="mean")
knn_median = KNN(n_neighbors=10, method="median")

调参：距离与方法

ps = [1, 2, 3, 4]
methods = ["largest", "mean", "median"]
scores = dict()


for p, method in product(ps, methods):
    # 初始化 KNN 实例
    knn = KNN(n_neighbors=10, method=method, p=p, n_jobs=-1)

    # 用当前 KNN 找到内点
    inliers = evaluate_outlier_classifier(knn, males_transformed, .55)

    # 计算 RMSE 并存入 scores
    scores[(p, method)] = evaluate_regressor(inliers)

结果检查

print(scores)

{(1, 'largest'): 23.188,
 (1, 'mean'): 23.188,
 (1, 'median'): 23.188,
 (2, 'largest'): 17.965,
 (2, 'mean'): 19.463,
 (2, 'median'): 19.463,
 (3, 'largest'): 17.965,
 (3, 'mean'): 19.463,
 (3, 'median'): 19.463, ...
}

Vamos praticar!

Python 中的异常检测