Implementando DBSCAN

Clase 16 de 27 • Curso de Clustering con Python y scikit-learn

Clase anteriorSiguiente clase

Alvaro Alva

student•

Que significa exactamente los parametros 'n_samples', 'noise', y 'random_state'?

Platzi

student•

Los parámetros 'n_samples', 'noise' y 'random_state' son utilizados en el contexto de la generación de datos aleatorios. 'n_samples' indica la cantidad de muestras o datos que se generarán. 'noise' se refiere al nivel de ruido o variabilidad que se añadirá a los datos generados. 'random_state' es una semilla que se utiliza para inicializar el generador de números aleatorios y asegurar que los resultados sean reproducibles.

Cristian perez

student•

en 2d y 3d se puede visualizar para saber la forma de los datos y asi escoger k means o db scan pero cuando son mas dimensiones como se haria para saber cual es mejor ya que no puedo visualizar las formas?

Octavio De Paula

student•

Es impresionante como ajustando los hiperparámetros se pueden identificar perfectamente tanto los "k" como los outliers.

dbscan_blobs = DBSCAN(eps=0.8, min_samples=4)
y_predict = dbscan_blobs.fit_predict(X)
df_blobs['cluster'] = y_predict
sns.scatterplot(data=df_blobs, x='x1', y='x2', hue='cluster', palette='bright');

Mario Alexander Vargas Celis

student•

Aquí tienes una implementación práctica de DBSCAN en Python usando scikit-learn, con visualización incluida:

✅ Paso 1: Cargar y preparar los datos

Usaremos datos simulados con formas no circulares para mostrar la ventaja de DBSCAN.

from sklearn.datasets import make_moons import matplotlib.pyplot as plt

# Crear un conjunto de datos de ejemplo (dos medias lunas) X, y = make_moons(n_samples=300, noise=0.1, random_state=42)

# Visualizar plt.scatter(X[:, 0], X[:, 1]) plt.title("Datos de ejemplo: dos medias lunas") plt.show()

✅ Paso 2: Aplicar DBSCAN

from sklearn.cluster import DBSCAN

# Crear modelo DBSCAN dbscan = DBSCAN(eps=0.3, min_samples=5) # Puedes ajustar estos parámetros

# Ajustar modelo y_dbscan = dbscan.fit_predict(X)

# Visualizar resultado plt.scatter(X[:, 0], X[:, 1], c=y_dbscan, cmap='plasma') plt.title("Clústeres encontrados con DBSCAN") plt.show()

⚙️ Parámetros importantes:

eps: distancia máxima entre dos puntos para que uno sea considerado vecino del otro.
min_samples: número mínimo de puntos para formar un clúster denso (incluye el punto central).

✅ Paso 3 (opcional): Identificar outliers

# Los puntos con etiqueta -1 son considerados ruido import numpy as np

outliers = X[y_dbscan == -1]

plt.scatter(X[:, 0], X[:, 1], c=y_dbscan, cmap='plasma') plt.scatter(outliers[:, 0], outliers[:, 1], c='red', label='Outliers') plt.legend() plt.title("DBSCAN con detección de outliers") plt.show()

Victor Jose Olivero

student•

Por que no se usa escalamiento para preparar los datos?

Jesús Ángel Martínez Briones

student•

Antonio Demarco Bonino

student•

Creo que me quedo con K-means

Juan José Mamani Tarqui

student•

<u>What is noise parameter?</u>

X_m ,y_m = make_moons(n_samples=250, noise=0.5, random_state=42)

In the context of machine learning and synthetic data generation, the term "noise" typically refers to random variations or perturbations added to the data. It introduces random fluctuations or errors to make the dataset more realistic and to simulate the inherent variability in real-world data.

In the specific case of the make_moons function from scikit-learn, the noise parameter controls the amount of random variation to be added to the data. The function generates a synthetic dataset representing two interleaving half circles, and the noise parameter allows you to control the level of random noise applied to the data points.

Here's an explanation of the parameters used in your code snippet:

n_samples: The total number of data points to generate.
noise: The standard deviation of the Gaussian noise added to the data. A higher value of noise will result in more scattered and less structured data points.
random_state: Seed for random number generation to ensure reproducibility.

So, in the case of make_moons(n_samples=250, noise=0.05, random_state=42), it generates a dataset with 250 samples representing two half circles, and each data point has some level of random noise (standard deviation of 0.05) added to its position. This is useful for creating a more realistic and challenging dataset for certain types of machine learning tasks, especially those dealing with non-linear decision boundaries.

Juan José Mamani Tarqui

student•

<u>Which would be the best noise?</u>

There isn't a single "best" noise level for the moons data. The ideal noise level depends on what you're trying to achieve:

Easy clustering task and baseline performance: If you want to establish a baseline performance for your clustering algorithm or want a very easy task to test with, then using a low noise level (e.g., noise=0.01 or even no noise) might be suitable.
Simulating real-world data: Real-world data often has some level of noise or uncertainty. Using a moderate noise level (e.g., noise=0.05 as you used) can be a good choice to make the clustering task more realistic and reflect the challenges of working with actual data.
Evaluating robustness to noise: If you're interested in testing how well your clustering algorithm handles noise, you might experiment with different noise levels (e.g., noise=0.05, 0.1, 0.2) and see how the performance changes. This can help you assess the algorithm's robustness.

Here are some additional factors to consider:

The clustering algorithm you're using: Some algorithms might be more sensitive to noise than others.
The desired level of difficulty: Do you want a clear separation for easy evaluation, or a more challenging scenario with some ambiguity?

Ultimately, the best noise level depends on your specific goals and the context of your work. It's a good practice to try different noise levels and see how they affect the clustering results.

Julián Cárdenas

student•

Buena clase sencilla y super explicativa !!

JOSE DANIEL ALVEAR ACEVEDO

student•

De las mejores explicaciones que he encontrado sobre el uso de DBSCAN

Julián Cárdenas

student•

Totalmente de acuerdo