Clustering Performance Evaluation on Different Preprocessing Techniques

This report analyzes the performance of three clustering algorithms (KMeans, Hierarchical, and Mean Shift) on a dataset using different preprocessing techniques. The evaluation metrics used include Silhouette Score, Calinski-Harabasz Index, and Davies-Bouldin Index. Each experiment was conducted for cluster sizes c = 3, 4, and 5.

---

📊 Clustering Performance Tables

✅ KMeans Clustering

Parameters	No Data Processing			Using Normalization			Using Standardization			Using PCA			Using T+N			T+N+PCA
	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5
Silhouette	0.5528	0.4981	0.4887	0.5048	0.4451	0.3526	0.4599	0.3869	0.3419	0.5092	0.4413	0.4156	0.4906	0.4392	0.3290	0.5508	0.5195	0.4291
Calinski-Harabasz	561.63	530.77	495.54	359.85	314.47	289.51	241.90	207.27	203.27	293.86	263.81	278.55	439.85	364.36	342.73	574.51	509.11	526.96
Davies-Bouldin	0.6620	0.7803	0.8060	0.7603	0.9004	0.9570	0.8336	0.8698	0.9530	0.7099	0.7560	0.7712	0.7760	0.9033	1.0211	0.6468	0.6954	0.7802

---

🧩 Hierarchical Clustering

Parameters	No Data Processing			Using Normalization			Using Standardization			Using PCA			Using T+N			T+N+PCA
	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5
Silhouette	0.5543	0.4890	0.4844	0.5048	0.4330	0.3493	0.4467	0.4006	0.3306	0.5111	0.4487	0.4042	0.4766	0.4223	0.3176	0.5330	0.5025	0.4089
Calinski-Harabasz	558.06	515.08	488.48	349.25	301.10	272.02	222.72	201.25	192.68	286.33	254.09	255.00	406.37	347.70	323.14	504.33	479.35	476.46
Davies-Bouldin	0.6563	0.7953	0.8204	0.7480	0.8486	0.9058	0.8035	0.9788	0.9742	0.7054	0.7226	0.7913	0.7697	0.8827	1.0020	0.5927	0.6240	0.7217

---

🌀 Mean Shift Clustering

Parameters	No Data Processing			Using Normalization			Using Standardization			Using PCA			Using T+N			T+N+PCA
	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5	c=3	c=4	c=5
Silhouette	NaN	NaN	NaN	0.4770	NaN	NaN	0.4325	NaN	NaN	0.4852	NaN	NaN	0.5444	NaN	NaN	NaN	0.4236	NaN
Calinski-Harabasz	NaN	NaN	NaN	290.47	NaN	NaN	163.71	NaN	NaN	167.41	NaN	NaN	278.23	NaN	NaN	NaN	255.82	NaN
Davies-Bouldin	NaN	NaN	NaN	0.7616	NaN	NaN	0.8131	NaN	NaN	0.6060	NaN	NaN	0.6702	NaN	NaN	NaN	0.7447	NaN

---

📈 Visualization

Clustering Performance Charts - KMeans, Hierarchical, Mean Shift

---

🔍 Observations

• KMeans with T+N+PCA preprocessing performed best overall with high Silhouette Score (0.5508) and low Davies-Bouldin Index (0.6468).
• Hierarchical clustering showed consistent results with PCA and T+N+PCA, especially for c=3 clusters.
• Mean Shift clustering had fewer valid results due to clustering instability or insufficient cluster separation, but preprocessing significantly improved performance.
• Using PCA and combined preprocessing (T+N+PCA) helps in improving clustering compactness and separation across methods.

---

✅ Conclusion

From the experimental analysis:

• T+N+PCA preprocessing is consistently beneficial for most clustering techniques.
• KMeans and Hierarchical clustering outperform Mean Shift in terms of available metrics and consistent results.
• PCA alone also provides better separation and compactness for some clustering methods.

This analysis provides clear guidance on choosing preprocessing methods to enhance clustering performance for downstream applications.