Dimensionality reduction visualization in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - Dimensionality reduction visualization

Which metric matters for Dimensionality Reduction Visualization and WHY

When we reduce data dimensions to show it in 2D or 3D, we want to keep the important structure. Metrics like Reconstruction Error tell us how much information is lost. Trustworthiness and Continuity measure if neighbors in the original data stay neighbors after reduction. These metrics help us know if the visualization truly represents the original data.

Confusion Matrix or Equivalent Visualization

Dimensionality reduction does not use a confusion matrix because it is not a classification task. Instead, we use Neighbor Preservation Tables or Distance Correlation to check if points close in original space remain close in reduced space.

Original neighbors: A B C D E
Reduced neighbors:  A B D C E

Neighbor preservation: 4 out of 5 neighbors kept
Trustworthiness = 0.8
    

Precision vs Recall Tradeoff Equivalent

In visualization, the tradeoff is between Preserving Local Structure (keeping close points close) and Global Structure (keeping overall shape). For example, t-SNE focuses on local neighbors (high trustworthiness) but may distort global distances. PCA keeps global variance but may lose local clusters. Choosing depends on what you want to see.

What "Good" vs "Bad" Metric Values Look Like

Good: Reconstruction error close to 0 means little info lost. Trustworthiness and continuity near 1 mean neighbors are well preserved. Visualization shows clear clusters or patterns matching original data.

Bad: High reconstruction error means much info lost. Trustworthiness below 0.5 means neighbors are not preserved well. Visualization looks messy or clusters are mixed up.

Common Pitfalls in Metrics

• Using only visual appeal without metrics can mislead about quality.
• Ignoring scale differences can distort distance-based metrics.
• Overfitting to training data can make visualization look good but fail on new data.
• Confusing global variance preservation with local neighbor preservation.

Self-Check Question

Your dimensionality reduction shows a reconstruction error of 0.3 (on scale 0 to 1) and trustworthiness of 0.4. Is this a good visualization? Why or why not?

Answer: No, because the reconstruction error is relatively high, meaning much information is lost. Trustworthiness is low, so neighbors are not well preserved. The visualization likely does not represent the original data well.

Key Result

Trustworthiness and reconstruction error are key metrics to evaluate how well dimensionality reduction preserves data structure for visualization.

Practice

1. What is the main purpose of dimensionality reduction in data visualization?

easy

A. To simplify complex data into fewer dimensions for easier understanding

B. To increase the number of features in the dataset

C. To remove all data points that are outliers

D. To convert numerical data into text format

Solution

1. Step 1: Understand dimensionality reduction goal

   Dimensionality reduction aims to reduce many features into fewer ones while keeping important information.

2. Step 2: Connect to visualization purpose

   Reducing dimensions helps us visualize data in 2D or 3D, making it easier to see patterns.

3. Final Answer:

   To simplify complex data into fewer dimensions for easier understanding -> Option A

4. Quick Check:

   Dimensionality reduction = simplify data [OK]

Hint: Dimensionality reduction means fewer features to see data clearly [OK]

Common Mistakes:

• Thinking it increases features instead of reducing
• Confusing it with data cleaning or outlier removal
• Believing it changes data type to text

2. Which of the following is the correct way to import PCA from scikit-learn in Python?

easy

A. import PCA from sklearn.decomposition

B. from sklearn.decomposition import PCA

C. from sklearn import PCA

D. import decomposition.PCA from sklearn

Solution

1. Step 1: Recall correct Python import syntax

   Python imports use 'from module import class' format for specific classes.

2. Step 2: Match with scikit-learn PCA import

   PCA is in sklearn.decomposition, so correct import is 'from sklearn.decomposition import PCA'.

3. Final Answer:

   from sklearn.decomposition import PCA -> Option B

4. Quick Check:

   Correct import syntax = from sklearn.decomposition import PCA [OK]

Hint: Use 'from module import class' for specific imports [OK]

Common Mistakes:

• Using wrong import order or syntax
• Trying to import PCA directly from sklearn
• Using 'import' with 'from' incorrectly

3. Given this Python code using PCA, what will be the shape of X_pca?

from sklearn.decomposition import PCA
import numpy as np
X = np.random.rand(100, 10)
pca = PCA(n_components=2)
X_pca = pca.fit_transform(X)

medium

A. (100, 2)

B. (2, 10)

C. (100, 10)

D. (10, 2)

Solution

1. Step 1: Understand input data shape

   X has 100 samples and 10 features, so shape is (100, 10).

2. Step 2: Apply PCA with n_components=2

   PCA reduces features from 10 to 2, keeping sample count same, so output shape is (100, 2).

3. Final Answer:

   (100, 2) -> Option A

4. Quick Check:

   Samples stay same, features reduced = (100, 2) [OK]

Hint: Output shape = (samples, n_components) after PCA [OK]

Common Mistakes:

• Confusing samples and features dimensions
• Thinking PCA changes number of samples
• Mixing up rows and columns in shape

4. This code tries to visualize data with t-SNE but throws an error. What is the likely cause?

from sklearn.manifold import TSNE
import numpy as np
X = np.random.rand(50, 5)
tsne = TSNE(n_components=3)
X_embedded = tsne.fit_transform(X)

medium

A. t-SNE does not support n_components greater than 2

B. t-SNE requires integer input data

C. Input data must have at least 10 features

D. t-SNE supports n_components=3, so no error here

Solution

1. Step 1: Check t-SNE n_components support

   t-SNE supports 2 or 3 components for visualization, so n_components=3 is valid.

2. Step 2: Verify input data and types

   Input X is float data with shape (50, 5), which is acceptable for t-SNE.

3. Final Answer:

   t-SNE supports n_components=3, so no error here -> Option D

4. Quick Check:

   t-SNE allows 3D output [OK]

Hint: t-SNE supports 2D or 3D outputs [OK]

Common Mistakes:

• Assuming t-SNE only supports 2D
• Thinking input must have many features
• Believing t-SNE needs integer data

5. You have a dataset with 1000 features and want to visualize clusters clearly. Which approach is best?

hard

A. Use PCA with 2 components because it is fast but may miss clusters

B. Use raw data without reduction for best cluster visualization

C. Use t-SNE with 2 components because it shows clusters well despite being slower

D. Use PCA with 100 components to keep most information

Solution

1. Step 1: Consider dataset size and visualization goal

   With 1000 features, dimensionality reduction is needed to visualize clusters clearly.

2. Step 2: Compare PCA and t-SNE for cluster visualization

   PCA is fast but focuses on variance, may miss clusters. t-SNE is slower but better at showing clusters in 2D.

3. Step 3: Choose best method for clear cluster visualization

   t-SNE with 2 components is best to reveal clusters clearly despite slower speed.

4. Final Answer:

   Use t-SNE with 2 components because it shows clusters well despite being slower -> Option C

5. Quick Check:

   t-SNE = better cluster visualization [OK]

Hint: t-SNE reveals clusters better than PCA in 2D [OK]

Common Mistakes:

• Choosing PCA for cluster clarity over t-SNE
• Using too many PCA components for visualization
• Trying to visualize raw high-dimensional data directly