Hyperparameters of KNN

Anomaly Detection in Python

Bekhruz (Bex) Tuychiev

Kaggle Master, Data Science Content Creator

Modify 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
Anomaly Detection in Python

Modifying 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 = mean_squared_error(y_test, preds, squared=False)

    return round(rmse, 3)
Anomaly Detection in Python

Tuning the number of neighbors

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


for k in n_neighbors:
    # Init a KNN
    knn = KNN(n_neighbors=k)

    # Get the inliers with KNN
    inliers = evaluate_outlier_classifier(knn, males_transformed, .55)

    # Calculate and store RMSE
    scores[k] = evaluate_regressor(inliers)
Anomaly Detection in Python

Inspecting the result

print(scores)

{5: 19.463, 10: 17.965, 15: 18.817, 20: 20.597}
Anomaly Detection in Python

Distance metrics

  • More than 40 metrics are supported
knn = KNN(metric='euclidean')

A plot of a sample dataset with 8 instances, with A as an outlier and arrows between A and its 4 closest neighbors.

Anomaly Detection in Python

Manhattan distance

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

573

The formula to calculate manhattan distance and a visual to illustrate the process.

Anomaly Detection in Python

Manhattan distance

  • Works well with high-dimensional data
  • Returns larger values than euclidean
  • Works best with categorical features

The formula to calculate manhattan distance and a visual to illustrate the process.

Anomaly Detection in Python

Minkowski distance

# Euclidean distance
knn = KNN(metric="minkowski", p=2)

# Manhattan distance
knn = KNN(metric="minkowski", p=1)

The formula to calculate minkowski distance and a visual to illustrate the process.

Anomaly Detection in Python

Distance aggregation

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


knn_mean = KNN(n_neighbors=10, method="mean")
knn_median = KNN(n_neighbors=10, method="median")
Anomaly Detection in Python

Tuning distance and method

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


for p, method in product(ps, methods):
    # Init an instance of KNN
    knn = KNN(n_neighbors=10, method=method, p=p, n_jobs=-1)

    # Find the inliers with the current KNN
    inliers = evaluate_outlier_classifier(knn, males_transformed, .55)

    # Calculate and store RMSE into scores
    scores[(p, method)] = evaluate_regressor(inliers)
Anomaly Detection in Python

Inspecting the result

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, ...
}
Anomaly Detection in Python

Let's practice!

Anomaly Detection in Python

Preparing Video For Download...