lucivpav / dnbc-scala

Parallel implementation of dynamic naive Bayesian classifier

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

dnbc-scala

Parallel implementation of dynamic naive Bayesian classifier

Download the paper.

Accuracy

Data sets based on Toy Robot data set

Data set type Average success rate [%]
Discrete 65
Continuous 42
Bivariate 76
Gaussian mixture (without hint) 96
Gaussian mixture (with hint) 99

The average success rate means the average percentage of hidden states inferred correctly.

There are two main reasons for relatively low overall sucess rate:

  1. Only about 90% of observed symbols are accurate
  2. There are multiple transitions to hidden states with the same observed symbol

Performance

Data set

Property Value
Number of hidden states 10
Sequence length 200
Observed discrete variables 5
Observed continuous variables 5
Learning set length (#sequences) 1000
Testing set length (#sequences) 200
Max Gaussians per mixture 3
Transitions per hidden state 5

Machine

Property Value
Processor 2× 8-core Intel Xeon E5-2650 v2 2.6 GHz
Memory 15 GB
Disk 10 GB HDD

Results

Property Workers=1 Workers=2 Workers=4 Workers=8 Workers=15
Learning time speed up 1 1.3 1.5 1.8 2

About

Parallel implementation of dynamic naive Bayesian classifier


Languages

Language:Java 75.7%Language:Scala 22.5%Language:Shell 1.9%