基于t-SNE降维坐标的高维汽车数据聚类可行性探讨

Great question! The short answer is yes, you absolutely can use t-SNE-reduced 2D coordinates for clustering, and this approach is indeed adopted by practitioners—but there are important caveats and best practices to address the instability issue you’ve noted.

Why t-SNE + Clustering Works #

t-SNE excels at preserving local data structure, meaning it does a far better job than PCA at highlighting tight, small clusters or fine-grained similarities in high-dimensional data. For your car dataset, this could help you identify nuanced groups (e.g., high-performance sports cars vs. fuel-efficient compact cars) that might get blurred by PCA or even direct clustering on raw features. Many data scientists and analysts use this combination when they want both actionable clusters and an intuitive visualization of how those clusters relate to each other.

Fixing t-SNE's Instability #

The randomness in t-SNE (from initialization and gradient descent steps) is a valid concern, but you can mitigate it with these steps:

• Fix the random seed: In libraries like scikit-learn, set random_state (e.g., TSNE(random_state=42)) when initializing the t-SNE model. This ensures the exact same 2D coordinates are generated every time you run it, making your clustering results reproducible.
• Optimize perplexity: Perplexity is a key parameter that balances local and global structure. For your 10,000-row dataset, try values between 20-30 (the default is 30) to get a more stable, meaningful embedding.
• Preprocess with PCA first: Most practitioners reduce the data to 50 dimensions via PCA before running t-SNE. This cuts down on noise, speeds up computation, and makes the t-SNE embedding more stable (since t-SNE struggles with extremely high raw dimensionality).
• Consensus clustering: If you want to account for minor random variations, run t-SNE multiple times with different seeds, cluster each embedding, then use consensus clustering to find cluster assignments that are consistent across runs.

Practical Recommendations #

• Validate clusters against raw features: t-SNE can distort global distances, so don’t rely solely on the 2D embedding’s clusters. Always check if the clusters make sense with your original features (e.g., do cluster 1 have consistently higher mpg than cluster 2? Are engine sizes distinct between groups?).
• Compare with other approaches: Run PCA+clustering and direct clustering (e.g., K-means on raw features) alongside t-SNE+clustering. Cross-reference the results to see which method produces the most interpretable, useful clusters for your use case.
• Leverage visualization: Pair your clustering results with the t-SNE plot—color-code points by cluster label to visually inspect how well-separated the groups are. This is one of the biggest perks of using t-SNE for clustering: you get both quantitative clusters and a qualitative view of their structure.

内容的提问来源于stack exchange，提问作者user3022875