MCEC algorithm

Program description

MCEC (Multivariate Correlations for Early Classification) algorithm is a Java implementation of an information-theoretic method for examining the early classification opportunity in a dataset. This dataset contains univariate or multivariate time series together with their respective class labels. The program can be downloaded here.

Input

The input file must be in comma-separated values (CSV) format, containing the time series and the respective class labels.

• Each line is expected to correspond to one instance.
• The columns must include the features grouped per time point, chronologically organized. The last column corresponds to the class attribute.
• The first line is the header, containing the names of the attributes and the associated time point. Example: "Att1_TP1", "Att2_TP1", "Att1_TP2", "Att2_TP2", "class".

Dataset example:

X1_1, X2_1, X1_2, X2_2, class
TRUE, FALSE, FALSE, FALSE, C1
FALSE, FALSE, TRUE, FALSE, C0
TRUE, TRUE, FALSE, FALSE, C0
TRUE, FALSE, TRUE, TRUE, C1
(...)

Output

The outcomes of the difference in entropy, log-likelihood, MDL score, AIC score and classification accuracy analysis, all for n = {1, ..., L} (where L represents the time series length) are outputted from the Java program in text files:

• dataEntropyDiff.txt.
• dataLL.txt.
• dataMDL.txt.
• dataAIC.txt.
• classResults.txt.

Observations

The proposed implementation provides the Markov Lag, an alternative to the standard Early Classification approach. Basically, instead of analysing the correlations from the initial time point until the last, it uses the inverse order (from the last to the first one). In this case, the idea is to check of how much information from the closest past we need, in order to obtain a satisfactory prediction.

Libraries

MCEC algorithm depends on two external libraries:

• Apache POI;
• Weka Data Mining Software.

Usage

Execute the jar file:

$ java -jar MCECalgorithm.jar dataset-filename.csv N optionClass MarkovLag

dataset-filename     Type: String - Name of the dataset file to be analysed.
N                    Type: Integer - Number of features per time point.
optionClass          Type: Boolean - With classification analysis (TRUE)
                     or without classification analysis (FALSE).
MarkovLag            Type: Boolean - With Markov lag approach (TRUE)
                     or with standard Early Classification (FALSE).

Synthetic dataset example

The very simple syntheticTest.csv dataset example is described in the following table:

The command for analysing the early classification opportunity is

$ java -jar MCECalgorithm.jar syntheticTest.csv 1 TRUE FALSE

and produces the following files:

• dataEntropyDiff.txt, with the difference in entropy results
• dataLL.csv, with the log-likelihood results
• dataMDL.txt, with the MDL score results
• dataAIC.txt, with the AIC score results
• classResultsL.txt, with the classification accuracy analysis