Lesson 8

Understanding and Comparing Clustering and Dimension Reduction Techniques


Today, we're turning our focus towards comparing various unsupervised learning methods. Our comparative study will include K-means, DBSCAN, Principal Component Analysis (PCA), Independent Component Analysis (ICA), and t-SNE.

Utilizing the Iris flower dataset, we will employ Python's scikit-learn library. Each of these methods possesses unique attributes, so understanding their comparative performance would enable us to choose the one best suited for any given scenario. Let's get started!

Understanding the Purpose of Comparison

In our exploration of unsupervised learning, we've familiarized ourselves with a variety of clustering and dimensionality reduction techniques. Although these techniques share the primary aim of discovering the underlying data structure, the methodologies they use to achieve this can vary significantly. That's where the need for comparison arises, as it helps us select the most suitable technique for a specific problem.

Several metrics, such as accuracy, simplicity, computational efficiency, and interpretability, enable us to compare these techniques. In the following sections, we'll compare clustering and dimension reduction methods using these metrics.

Comparing Clustering Techniques

Let's begin by refreshing our memory on the properties of our clustering techniques. K-means is a partition-based technique. It partitions observations into clusters in such a way that each observation belongs to the cluster with the nearest mean. The clusters formed by K-means tend to be spherical, which suits well-spaced, round clusters. However, it doesn't handle noise and outliers effectively and struggles with non-spherical and similarly sized clusters.

In contrast, DBSCAN is a density-based clustering algorithm. It considers clusters as dense regions separated by regions of lower density in the feature space — hence, it can capture clusters of arbitrary shapes, a clear advantage over K-means. Moreover, it can handle noise in the data. However, deciding appropriate parameters such as eps and min_samples can be a bit tricky, and this algorithm may struggle with clusters of differing densities.

K-Means vs. DBSCAN

K-means and DBSCAN, as clustering techniques, can be compared across several parameters:

  • Cluster Quality: K-means excels in creating spherical and similarly sized clusters, while DBSCAN outperforms it in forming clusters of varying shapes and sizes.
  • Scalability: While K-means easily scales with large datasets, DBSCAN often requires additional computational resources as the dimensions increase.
  • Tolerance to Noise: DBSCAN identifies and handles noise and outliers effectively, giving it an advantage over K-means, which often absorb noisy points into clusters.
  • Working with Different Density: DBSCAN adapts well according to the density-based definition of a cluster, while K-means might struggle with clusters of varying densities.
  • Interpretability: K-means provides intuitive and easy-to-interpret results, while DBSCAN’s results may be slightly harder to interpret.

The above comparison between K-means and DBSCAN will make it easier to decide which method meets your specific requirements. For instance, if your data contains noise or necessitates flexible cluster shapes, DBSCAN may offer a more suitable choice.

Comparing Dimension Reduction Techniques

Next, let's talk about dimensionality reduction techniques.

Principal Component Analysis (PCA), Independent Component Analysis (ICA), and t-Distributed Stochastic Neighbor Embedding (t-SNE) are all statistical techniques used for dimensionality reduction. They all have strengths and weaknesses and can be best applied based on the specifics of the given dataset.

PCA is an unsupervised method that is most useful in an exploratory scenario where we're not quite sure what we're looking for. PCA aims to find the directions (principal components) that maximize the variance of the data. It assumes that data axes with greater variance are more significant. However, these axes might not be optimal for separating different classes in the data, so it is possible that PCA could end up removing the features that are critical for discrimination.

On the contrary, ICA is also an unsupervised method, but while PCA focuses on maximizing variance, ICA aims to find the statistically independent axes. Essentially, it assumes that the observed data are linear combinations of some unknown independent components. This makes ICA especially useful when you need to separate mixed signals, such as separating out background noise from music. However, ICA assumes non-Gaussian statistics, which may not always be the case.

T-SNE, like PCA, is an unsupervised technique. It computes the probability that pairs of data points in the high-dimensional space are related and then chooses a low-dimensional embedding that produces a similar distribution. Unlike PCA and ICA, t-SNE does not suggest a linear mapping from high- and low-dimensional space. Hence, it can capture complex polynomial relationships between features. It is particularly good at preserving local structure, making it excellent for exploratory data analysis. But this strength is also its weakness. Preserving local structure often comes at the expense of distorting global structure. The degree of distortion increases with increasing number of dimensions. Also, it tends to be computationally expensive, especially with large datasets.

Additionally, t-SNE has a few hyperparameters like perplexity and learning rate that can significantly affect the output, and it's not always clear how to choose these.

In summary, the choice of PCA, ICA, or t-SNE really depends on the specific dataset and the problem you are trying to solve. PCA is good if you think a linear model could describe your data. ICA is good if your data is thought to comprise independent components. Lastly, t-SNE is a great exploratory tool if you're working with high-dimensional data and want to visualize it in a low-dimensional space.

PCA vs. ICA vs. t-SNE

These techniques can be distinguished based on the following criteria:

  • Explained Variance: PCA directly measures the retained variance in the transformed data, while ICA and t-SNE do not offer as explicit a measure.
  • Computational Efficiency: PCA demands fewer computing resources than ICA and t-SNE.
  • Interpretability: PCA and ICA produce encoded dimensions that are interpretable, unlike the reduced dimensions in t-SNE, which aren't directly interpretable.
  • Modelling Technique: While all three function as unsupervised techniques and are independent of any labels, they aim to achieve different things. PCA looks for the greatest variance, ICA looks for statistical independence, while t-SNE makes probability distributions in different dimensions as similar as possible.
Comparative Analysis on Iris Dataset

Now, let's move to the exciting part: applying the clustering and dimensionality reduction techniques to the Iris dataset and extrapolating insights.

For instance, we might find that K-means and PCA perform well together because K-means leverage PCA's efficiency and interpretability when reducing data dimensions.

Here's a concise snippet of code exemplifying the application of K-means and PCA together, visually displaying the results for better understanding.

1# Import required libraries 2from sklearn.cluster import KMeans 3from sklearn.decomposition import PCA 4from sklearn import datasets 5import matplotlib.pyplot as plt 6 7# Load Iris dataset 8iris = datasets.load_iris() 9 10# Apply PCA to reduce dimensions to 2 11pca = PCA(n_components=2).fit_transform(iris.data) 12 13# Apply K-means with 3 clusters, matching the number of iris species 14km = KMeans(n_clusters=3) 15km.fit(pca) 16 17# Plotting PCA and K-means results 18plt.scatter(pca[:, 0], pca[:, 1], c=km.labels_) 19plt.title('Clustering Iris Dataset using K-means and PCA') 20plt.show()


The code above provides us with a visual representation of the Iris data, divided into three distinct clusters. These clusters are formed by applying K-means to the data that has been reduced to two dimensions using PCA. Each point on the plot represents an Iris flower, color-coded based on its predicted cluster. The clarity between clusters suggests that K-means has successfully classified the transformed data.

Analyzing this comparative breakdown of methods on the same dataset provides a clearer understanding of when to employ a specific technique and what to expect from it.


Congratulations on making it this far! We've meticulously compared various clustering and dimensionality reduction techniques to elucidate the strengths and limitations of each. This knowledge will enable you to make an informed choice when confronted with a new dataset or problem.

Not only do these comparisons give an understanding of the capabilities of each method, but they also recall our progress thus far in the course. Additionally, they stoke our curiosity and learning desire, paving the path we'll be taking in the forthcoming lessons. As we move ahead, understanding the strengths and weaknesses of these unsupervised learning methods will aid our transition into a more advanced stage of machine learning.

Enjoy this lesson? Now it's time to practice with Cosmo!

Practice is how you turn knowledge into actual skills.