BLADE

---

Overview

The study of multivariate time-series is applied to a wide range of areas in society, being used to represent all sorts of events and behaviors across time. The BLADE (Boltzmann dynamic network) model aggregates both a dynamic Bayesian networks approach to study time-series data and a restricted Boltzmann machine implementation to derive hidden feature relationships present along the time frame.

Assumptions

• In this implementation of BLADE, the data is considered to be categorical (features assume one of multiple finite values). Some method of discretization such as symbolic aggregate approximation (SAX) can be used to discretize a real-valued dataset. Some examples of this case were used and are provided in the directory real_data

How to run the RBM-tDBN algorithm

1. Download a Python distribution

2. Download the repository

git clone https://github.com/joaor96/rbm-tdbn
cd rbm-tdbn

3. If using python >= 3.6:

   1. Add the root directory to the PYTHONPATH

      export PYTHONPATH="${PYTHONPATH}:path/to/BLADE"
      

   2. Create a virtual environment with the packages required, that can be found on the file required_packages.txt

4. If using Anaconda, create a virtual environment with the packages required, that can be found on the file required_packages.txt

5. The runnning mode flag that is present in the file blade.py defines the running mode, with two different possible values:

   • "dev": used when running the code with the values defined on the file, for testing purposes.
   • "run": used when running the code through the command line, with the parameter values indicated by the user.

6. If running in terminal, input the following:

   blade.py <filepath> <options>
   

Parameters

The different parameters that can be specified when running the BLADE algorithm are presented in the table below.

parameter	parameter code		default	description
filepath		✔️	--	file path for both the data file (filepath+``\_parsed.csv'') and the labels file (filepath + ``\_labels.csv'')
tdbn\_parents	-tdbnp		1	set the number of parent nodes to be considered by the \acrshort{tdbn} algorithm
no\_rbm	-nrbm		False	if the user wishes to run the method without the \acrshort{rbm} pre-processing, he should indicate it by placing this parameter on the input
test\_set\_ratio	-tsr		0.2	ratio of the original data to be used for testing
validation\_set\_ratio	-vsr		0.2	ratio of the pre-training dataset to be used for validation
batch\_size\_ratio	-bsr		0.1	ratio to be used when splitting the training set in mini-batches for the \acrshort{cd} algorithm
hidden\_units	-hu		3	number of hidden units to be used in the \acrshort{rbm}s
epochs	-e		100	defines the number of iterations in the \acrshort{cd}-1 algorithm
learning\_rate	-lr		0.05	learning rate for the \acrshort{rbm}
weight\_decay	-wd		1x10-4	weight decay value when training the \acrshort{rbm}
persistent\_cd	-pcd		False	defines the usage of persistent \acrshort{cd}.
number\_runs	-nr		10	since the accuracy has a variance associated, increasing the number of runs of the algorithm ensures that a better approximation of the actual accuracy value is calculated
validation\_runs	-vr		5	number of validation runs
extraction\_runs	-er		5	number of extractions in each validation cycle
verbose	-vb		False	if \texttt{True}, it prints the results of training the \acrshort{rbm} and \acrshort{tdbn} while the \acrshort{blade} algorithm is running
version	-v		--	prints the current version of the \acrshort{blade} implementation

Future work

References

• tDBN algorithm, extended to create an initial network and calculate the probabilities of a test dataset, developed by José Monteiro
• RBM implementation in Python using Pytorch, used as a basis for the implementation here present, and developed by Gabriel Bianconi, under the MIT License.