Distance-based Features

[illuminoplanet]

Align the failure points according to their Euclidean distance from the center. Then extract following features:

1. Frequency by the interval of distance, 10 bins were made but was cubic interpolated to extract 20 features
2. Mean
3. Standard Deviation
4. Max
5. Min
6. Arg max
7. Arg min

Altogether 26 features were extracted.

[illuminoplanet]

Implementation

def extract_distance(x):

    feature_name = lambda s, x: f"{s}_{str(x).zfill(2)}"

    coor = np.argwhere(x==2)-(np.array(x.shape)//2)
    radius = np.linalg.norm(coor, ord=2, axis=1)

    dist = {}

    # polar accumulate 
    dist_y, _ = np.histogram(radius)
    dist_x = np.linspace(1, dist_y.size, dist_y.size)

    dist_interpolate = interp1d(dist_x, dist_y, kind='cubic')
    new_dist_x = np.linspace(1, dist_y.size, 20)
    new_dist_y = dist_interpolate(new_dist_x)/np.linspace(1, new_dist_x.size, new_dist_x.size)
    
    for i in range(20):
        dist[feature_name('dist_value', i+1)] = new_dist_y[i]

    dist[feature_name('dist', 'mean')] = np.mean(new_dist_y)
    dist[feature_name('dist', 'std')] = np.std(new_dist_y)
    dist[feature_name('dist', 'max')] = np.max(new_dist_y)
    dist[feature_name('dist', 'min')] = np.min(new_dist_y)
    dist[feature_name('dist', 'argmax')] = np.argmax(new_dist_y)
    dist[feature_name('dist', 'argmin')] = np.argmin(new_dist_y)

    return pd.Series(dist)

[illuminoplanet]

Evaluation

Control: Density-based + Radon-based + Geometry-based
Estimator 1 (LR) :
          Accuracy : 62.51%
          AUC : 0.9142
Estimator 2 (RF) :
          Accuracy : 80.08%
          AUC : 0.9752
Estimator 3 (GBM) :
          Accuracy : 79.53%
          AUC : 0.9708
Estimator 4 (ANN) :
          Accuracy : 70.12%
          AUC : 0.9402

Experiment: Density-based + Radon-based + Geometry-based + *Distance-based
Estimator 1 (LR) :
          Accuracy : 68.55%
          AUC : 0.9355
Estimator 2 (RF) :
          Accuracy : 81.73%
          AUC : 0.9771
Estimator 3 (GBM) :
          Accuracy : 81.57%
          AUC : 0.9781
Estimator 4 (ANN) :
          Accuracy : 71.92%
          AUC : 0.9475

[illuminoplanet]

Analysis

Showed minor improvement in all estimators and major improvement on the logistic regression classifier.

[ssupecial]

제 결과표(Evaluation)에 위의 Evaluation을 포함하여 적도록 하겠습니다