EXP2: Kmeans

[WeiFoo]

Technical Details.

• Cluster Testing data, has centroids A, B, C....., the average distance between data and centroids is d .
• Calculate distance between each tuning data and the nearest cluster centroids.
• Collect the tuning data if their distance to nearest cluster centroids <= **d***1.25 (engineering judgment)
• Return those data as final tuning data.

  number of clusters = 3

  F

Precision

number of clusters = 5

F

Precision

[WeiFoo]

Notes:

UP to now, I can't see an exciting method that can improve the performance of tuning, neither K-means nor DBSCAN.

Next, I will try WHERE.

[timm]

one thing you can do with where is split at median point

anything you can do is apply sdiv to split at points that reduces the variation in distances between instances.