pythontime-seriesunsupervised-learning
Evaluation metric for parameter tuning for outlier detection (unsupervised learning) on time series
I'm working on implementing parameter tuning for outlier detection in time-series data using the DBSCAN algorithm. To maximize the Silhouette score (as evaluation), I'm leveraging optuna for tuning. However, after parameter tuning, the model's performance seems to be underperformed. Below is the complete code, which encompasses data generation, preprocessing, decomposition, parameter tuning, and applying.
I utilized isolated forest, LOF, and OneSVM algorithms and the result was similar. I utilized metrics including davies_bouldin_score and calinski_harabasz_score, but did not achieve better results.
How can I improve the outlier detection parameter tuning?
import numpy as np
import matplotlib.pyplot as plt
from statsmodels.tsa.seasonal import seasonal_decompose
from sklearn.preprocessing import MinMaxScaler
from sklearn.cluster import DBSCAN
from sklearn.metrics import silhouette_score
import optuna
# Function to generate time series data
def generate_time_series(n_samples=300, n_outliers=30):
np.random.seed(np.random.randint(10000))
t = np.linspace(0, 50, n_samples)
y = np.cumsum(np.random.randn(n_samples)) + np.sin(t) # Adding trend and noise
outlier_indices = np.random.choice(n_samples, n_outliers, replace=False)
y[outlier_indices] += 15 * np.random.randn(n_outliers) # Injecting outliers
return y.reshape(-1, 1), t
# Generate the time series data
y, t = generate_time_series()
# Plot the time series data
plt.figure(figsize=(10, 5))
plt.plot(t, y, label='Time series', color='blue')
plt.xlabel('Time')
plt.ylabel('Value')
plt.title('Generated Time Series Data')
plt.legend()
plt.show()
# Decompose the time series
result = seasonal_decompose(y, period=30, model='additive', two_sided=True)
residual = result.resid
# Handle NaN values in residuals (if any)
non_nan_indices = ~np.isnan(residual).flatten()
residual = residual[non_nan_indices].reshape(-1, 1)
t_residual = t[non_nan_indices]
# Plot the seasonal decomposition
plt.figure(figsize=(10, 5))
plt.subplot(411)
plt.plot(t, y, label='Original', color='blue')
plt.legend(loc='best')
plt.subplot(412)
plt.plot(t, result.trend, label='Trend', color='orange')
plt.legend(loc='best')
plt.subplot(413)
plt.plot(t, result.seasonal, label='Seasonal', color='green')
plt.legend(loc='best')
plt.subplot(414)
plt.plot(t_residual, residual, label='Residual', color='red')
plt.legend(loc='best')
plt.tight_layout()
plt.show()
# Scale the residual data
scaler = MinMaxScaler()
residual_scaled = scaler.fit_transform(residual)
# Define the objective function for DBSCAN
def dbscan_objective(trial):
eps = trial.suggest_float('eps', 0.01, 0.5, log=True)
min_samples = trial.suggest_int('min_samples', 2, 20)
dbscan = DBSCAN(eps=eps, min_samples=min_samples)
clusters = dbscan.fit_predict(residual_scaled)
# Ignore cases where all points are considered noise
if len(set(clusters)) <= 1:
return -1.0
score = silhouette_score(residual_scaled, clusters)
return score
# Optimize DBSCAN using Optuna
optuna.logging.set_verbosity(optuna.logging.WARNING)
dbscan_study = optuna.create_study(direction='maximize')
dbscan_study.optimize(dbscan_objective, n_trials=100, show_progress_bar=True)
best_dbscan_params = dbscan_study.best_params
print(f"Best DBSCAN parameters: {best_dbscan_params}")
# Apply DBSCAN with the best parameters
dbscan = DBSCAN(**best_dbscan_params)
dbscan_clusters = dbscan.fit_predict(residual_scaled)
dbscan_outliers = (dbscan_clusters == -1)
# Plot the detected outliers in the residuals
plt.figure(figsize=(10, 5))
plt.plot(t_residual, residual, label='Residual', color='blue')
plt.scatter(t_residual[dbscan_outliers], residual[dbscan_outliers], color='red', label='Outliers')
plt.xlabel('Time')
plt.ylabel('Value')
plt.title('DBSCAN Outlier Detection on Residuals')
plt.legend()
plt.show()
# Plot the detected outliers in the original time series
plt.figure(figsize=(10, 5))
plt.plot(t, y, label='Time series', color='blue')
plt.scatter(t_residual[dbscan_outliers], y[non_nan_indices][dbscan_outliers], color='red', label='Outliers')
plt.xlabel('Time')
plt.ylabel('Value')
plt.title('DBSCAN Outlier Detection on Original Time Series')
plt.legend()
plt.show()
# Print the number of outliers detected by DBSCAN
print(f"Number of outliers detected by DBSCAN: {np.sum(dbscan_outliers)}")
Solution
DBSCAN relies on distance measurements to find clusters, thus it is sensitive to the scale and distribution of the data. Even, in your case, you have just one feature vector, I don't think you need to scale it for outlier detection. Just use residual variable in hyper-parameter and final prediction. You may also need to increase eps may be up to 2. So final code would look like this:
import numpy as np
import matplotlib.pyplot as plt
from statsmodels.tsa.seasonal import seasonal_decompose
from sklearn.cluster import DBSCAN
from sklearn.metrics import silhouette_score
import optuna
# Function to generate time series data
def generate_time_series(n_samples=300, n_outliers=30):
np.random.seed(np.random.randint(10000))
t = np.linspace(0, 50, n_samples)
y = np.cumsum(np.random.randn(n_samples)) + np.sin(t) # Adding trend and noise
outlier_indices = np.random.choice(n_samples, n_outliers, replace=False)
y[outlier_indices] += 15 * np.random.randn(n_outliers) # Injecting outliers
return y.reshape(-1, 1), t
# Generate the time series data
y, t = generate_time_series()
# Plot the time series data
plt.figure(figsize=(10, 5))
plt.plot(t, y, label='Time series', color='blue')
plt.xlabel('Time')
plt.ylabel('Value')
plt.title('Generated Time Series Data')
plt.legend()
plt.show()
# Decompose the time series
result = seasonal_decompose(y, period=30, model='additive', two_sided=True)
residual = result.resid
# Handle NaN values in residuals (if any)
non_nan_indices = ~np.isnan(residual).flatten()
residual = residual[non_nan_indices].reshape(-1, 1)
t_residual = t[non_nan_indices]
# Plot the seasonal decomposition
plt.figure(figsize=(10, 5))
plt.subplot(411)
plt.plot(t, y, label='Original', color='blue')
plt.legend(loc='best')
plt.subplot(412)
plt.plot(t, result.trend, label='Trend', color='orange')
plt.legend(loc='best')
plt.subplot(413)
plt.plot(t, result.seasonal, label='Seasonal', color='green')
plt.legend(loc='best')
plt.subplot(414)
plt.plot(t_residual, residual, label='Residual', color='red')
plt.legend(loc='best')
plt.tight_layout()
plt.show()
# Define the objective function for DBSCAN
def dbscan_objective(trial):
eps = trial.suggest_float('eps', 0.01, 2, log=True)
min_samples = trial.suggest_int('min_samples', 2, 20)
dbscan = DBSCAN(eps=eps, min_samples=min_samples)
clusters = dbscan.fit_predict(residual)
# Ignore cases where all points are considered noise
if len(set(clusters)) <= 1:
return -1.0
score = silhouette_score(residual, clusters)
return score
# Optimize DBSCAN using Optuna
optuna.logging.set_verbosity(optuna.logging.WARNING)
dbscan_study = optuna.create_study(direction='maximize')
dbscan_study.optimize(dbscan_objective, n_trials=100, show_progress_bar=True)
best_dbscan_params = dbscan_study.best_params
print(f"Best DBSCAN parameters: {best_dbscan_params}")
# Apply DBSCAN with the best parameters
dbscan = DBSCAN(**best_dbscan_params)
dbscan_clusters = dbscan.fit_predict(residual)
dbscan_outliers = (dbscan_clusters == -1)
# Plot the detected outliers in the residuals
plt.figure(figsize=(10, 5))
plt.plot(t_residual, residual, label='Residual', color='blue')
plt.scatter(t_residual[dbscan_outliers], residual[dbscan_outliers], color='red', label='Outliers')
plt.xlabel('Time')
plt.ylabel('Value')
plt.title('DBSCAN Outlier Detection on Residuals')
plt.legend()
plt.show()
# Plot the detected outliers in the original time series
plt.figure(figsize=(10, 5))
plt.plot(t, y, label='Time series', color='blue')
plt.scatter(t_residual[dbscan_outliers], y[non_nan_indices][dbscan_outliers], color='red', label='Outliers')
plt.xlabel('Time')
plt.ylabel('Value')
plt.title('DBSCAN Outlier Detection on Original Time Series')
plt.legend()
plt.show()
# Print the number of outliers detected by DBSCAN
print(f"Number of outliers detected by DBSCAN: {np.sum(dbscan_outliers)}")
And you will get something like this:
- Implementation of Okapi BM25 in python
- How do I parse a string to a float or int?
- What is the best design pattern when creating an interface for functions that can return a variety of responses?
- How can I sandbox Python in pure Python?
- Currency formatting in Python
- How boxen plot is different from box plot?
- Sorting names by their high scores
- How do I serve both HTML and JSON content with Flask?
- Changing lowercase characters to uppercase and vice-versa in Python
- WARNING Not Found: /favicon.ico in Django
- SQL Alchemy - How to delete a model instance?
- What does the python operator =- do?
- How can I unit test the jinja2 template logic?
- python pandas dataframe select rows with ternary operator
- Selenium using Python - Geckodriver executable needs to be in PATH
- Gunicorn do not reload worker
- How to add timestamp to each request in uvicorn logs?
- How to extract video titles with the Youtube API(Python)
- Cartesian product of dict of lists in Python
- Round to 5 (or other number) in Python
- Why do I get a "ModuleNotFoundError" in VS Code despite the fact that I already installed the module?
- Python: Suppressing errors from going to commandline?
- Problems using MySQL with AWS Lambda in Python
- Is there any benefit of using a dictionary comprehension if an equivalent dict(zip(a, b)) is possible?
- SQLAlchemy - Get query results in same order as IN clause
- Python3 __pycache__ generating even if PYTHONDONTWRITEBYTECODE=1
- Matplotlib.animation: how to remove white margin
- bash script maybe syntax error not sure trying a until loop
- How to typehint dataclass field in function signature
- pdfium library is not working when compiling my code in python