Unsupervised Algorithms and Metrics

[brifordwylie]

𝐭 𝐡𝐨𝐰 𝐝𝐨 𝐰𝐞 𝐤𝐧𝐨𝐰 𝐰𝐡𝐢𝐜𝐡 𝐦𝐞𝐭𝐡𝐨𝐝 𝐢𝐬 𝐛𝐞𝐭𝐭𝐞𝐫? 𝐖𝐞 𝐝𝐨𝐧’𝐭 𝐡𝐚𝐯𝐞 𝐥𝐚𝐛𝐞𝐥𝐬 𝐢𝐧 𝐔𝐧𝐬𝐮𝐩𝐞𝐫𝐯𝐢𝐬𝐞𝐝 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠, 𝐍𝐨 𝐠𝐫𝐨𝐮𝐧𝐝 𝐭𝐫𝐮𝐭𝐡.

The answer lies in using evaluation metrics that can help us determine the quality of our algorithm.

𝔼𝕧𝕒𝕝𝕦𝕒𝕥𝕚𝕠𝕟 𝕄𝕖𝕥𝕙𝕠𝕕𝕤:

➊ Silhouette score:

A high Silhouette score (close to 1) indicates that data points within clusters are similar, and that the normal data points are well separated from the anomalous ones.

➋ Calinski-Harabasz index:

Calinski-Harabasz Index measures the between-cluster dispersion against within-cluster dispersion. A higher score signifies better-defined clusters.

➌ Davies-Bouldin index:

Davies-Bouldin Index measures the size of clusters against the average distance between clusters. A lower score signifies better-defined clusters.

➍ Kolmogorov-Smirnov statistic:

It measures the maximum difference between the cumulative distribution functions of the normal and anomalous data points.

➎ Precision at top-k:

The metric calculates the precision of the top-k anomalous data points using expert domain knowledge.

https://towardsdatascience.com/7-evaluation-metrics-for-clustering-algorithms-bdc537ff54d2

https://towardsdatascience.com/three-performance-evaluation-metrics-of-clustering-when-ground-truth-labels-are-not-available-ee08cb3ff4fb

https://medium.datadriveninvestor.com/evaluation-metrics-for-clustering-96dcdbea437d

https://towardsdatascience.com/a-comprehensive-beginners-guide-to-the-diverse-field-of-anomaly-detection-8c818d153995