Like-Behaviors Labeling Routing (LBLR)

LBLR presented in Efficient and Effective Labeling of Massive Time Series Archives (Submitted to KDD 2018) is our solution to the following problem:

Given a time series T of length n, we wish to produce an annotation vector A of length n which correctly
annotates T while minimizing the number of accesses to a human annotator.

Here you will find the project Source Code, Documentation, and Experimental Results but be sure to also checkout our project website LBLR Homepage.

Though not fully optimized we encourage users to use our application and share any experiences (or issues) they encounter when using our software. Please send any inquiries to fmadr002[at]ucr[dot]edu.

Getting Started

Prerequisites

This project requires an installation of Matlab (R2017b) and the Communications Toolbox.

Installing

After installing the required software and toolboxes, simply clone or download the project to your Matlab User Path. To identify your Matlab User Path, run the command userpath on the Matlab command line. By default on a Windows 7 or Windows 10 machine, the userpath is C:\Users\[username]\Documents\MATLAB.

To test your setup, simply run the application with the provided test input: LBLR

If the application initiated, then setup was successful.

Using LBLR

The LBLR application can be initiated by simply executing LBLR in the MATLAB command window. To execute AutoLBLR, call AutoLBLR(...) with the appropriate parameters (see below)

AutoLBLR is an unsupervised execution of the LBLR application. Instead of deferring to the human annotator (user) AutoLBLR will simply apply a classification label to the cluster of subsequences which are semantically similar. If AutoLBLR is not running blind it will greedily apply the most likely label as indicated by the SolutionVector otherwise the label will simply be the current iteration of the algorithm.

[Labels, PlotHandles, Completion] = AutoLBLR(TimeSeries, ModelModelLength, Solution = [], OPTIONS)

    Inputs:
        TimeSeries     - A numeric non-empty vector with length n
        ModelLength    - A scalar integer less than n
        SolutionVector - Classification labels for the corresponding data poitns in 'TimeSeries' representing the ground truth.
                         If empty, LBLR will be assumed to operate blind; otherwise, must be a non-empty numeric integer vector
                         with length n.

    Outputs:
        Labels      - Predicted labels for the corresponding points in 'TimeSeries'
        PlotHandles - An array of figure handles for each of the generated plots.
        Completion  - Increasing numeric vector indicating the percentage of annotated data after each iteration.

    Options
        ExclusionRange    - A heuristically set window which excludes a range of subsequences from the original subsequene when
                            checking for similarity.
        Bits              - Number of bits used to discretize the TimeSeries.
        Blind             - If enabled, AutoLBLR will NOT math predicted labels to a solution vector. This option is useful for
                            identifying preserved subsequences within a defined behavior.
        Debug             - Outputs a deeper trace to the output file. Used for internal purposes only or really curious users.
        MaximumIterations - Sets a limit to the amount of iterations AutoLBLR would run.

Built With

Matlab (R2017b)

Contributing

We are not accepting external commits at this time but will be facilitating the means for public contribution in the near future.

Authors

Frank Madrid - Initial work - fmadrid

See also the list of contributors who participated in this project.

franzbischoff / LBLR