Clustering data structure, DP prior and split-merge sampler

[ubc-stat-ml]

Clustering data structure, DP prior and split-merge sampler

[alexandrebouchard]

Start with just a simpler one point at time sampler - annealing migh be enough.

Need:

• Exch : both for large observed data and internal to DPs
  • Matrix getSuffStat(T => Matrix) + typical ones moments, sum of logs, etc [always additive, so can assume will return a matrix; wait to see if other types of suff stats are needed, probably not?]
  • separate observed vs unobserved for efficiency (and put @immutable on former) <- no, will complicate things for DP machinery
  • add and remove (remove needed for DP machinery); take cares of listeners, etc
  • parser, writer
• ExchIntVar, ExchRealVar, ExchMatrix
  • addSuffStatCollector
• SuffStatCollector has addToSuffStat
• Then need some distributions on Exch stuff, e.g. PoiGamma, etc. Internally, they just call another one e.g. NegBinomial with parameters extracted from the suffstat
• Then need clustering data structure
• Move: a one bit change thing. Argg at the end we pay too much by number of clusters with that. Oh well can create a sampler attached to a model. Still useful for DP vs PitmanYor vs discrete.
• Prior over clusterings
• ClusterLikelihood: exch<T> | cluster, Distrib<Exch>