Options for HDBSCAN

[dchiu911]

Here are my thoughts on incorporation @Dustin21:

1. Do not include with the rest of the algorithms for consensus consideration.
2. Use it as a metric for reporting "outliership" and number of clusters for each run of the algorithm.

I've already made a few changes:

• DBSCAN --> HDBSCAN: hierarchical structure spans over multiple eps values, so we only need to specify single paramter minPts
• minPts default: 5 --> 2

What do you think?

[Dustin21]

That's a great idea. I completely forgot about HDBSCAN. It also provides a soft clustering correct? Also, using it for outliership is exactly what I had in mind. What is your reason behind the default of 5?

[Dustin21]

@dchiu911 I think there may be a way to choose k with HDBSCAN during the hierarchical step. I will look into this.

[dchiu911]

I changed from 5 to 2 because it resulted in fewer noise points.

[dchiu911]

@Dustin21 so I made an initial implementation for HDBSCAN, code is here

Whenever it is called, the clustering is separated from the rest of the consensus. Instead, we report, for each replication, the number of clusters and the proportion of outliers.

[Dustin21]

It looks great! I'll give it a test run on Genesis next week with the parameter defaults.

[dchiu911]

Barring major errors, going to close. Reminder: dbscan package has

SystemRequirements: C++11

hence keeping it in dev version of diceR