The repository contains the winning solution to the AAIA'15 Data Mining Competition: Tagging Firefighter Activities at a Fire Scene.
R
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
data
doc
scripts
submissions
.gitignore
README.md

README.md

AAIA`15 Data Mining Competition: Tagging Firefighter Activities at a Fire Scene

Author

Jan Lasek (Institute of Computer Science, Polish Academy of Sciences), janek.lasek@gmail.com.

General info

The repository contains the winning solution to the AAIA`15 Data Mining Competition: Tagging Firefighter Activities at a Fire Scene organized by University of Warsaw, Poland and Main School of Fire Service, Warsaw Poland. The goal of the contest was to develop a classifier for tagging a firefighter's actions based on sensory recordings from accelerometers and gyroscopes attached to different parts of his/her body. Pay a visit at the competition's hosting platform for more information.

The short README file herein describes how to use the enclosed solution. The steps for feature extraction, model estimation and submission generation are presented. The experiments were performed on a single machine with Intel(R) Core(TM) i7-4510U CPU @ 2.00GHz and 16 GB RAM running Ubuntu 14.04.2 LTS. The solution is implemented in R, a language and environment for statistical computing.

Solution

Data

Data for running experiment should be downloaded from competition website. There are two datasets: trainingData.csv and testData.csv along with a file containing target labels trainingLabels.csv for the training set. Each dataset contains 20,000 instances of activities performed by different firefighters. Each dataset is of size 2.4 GB. The goal is to label the instances in the test set.

Feature extraction

Script feature_extraction.R in scripts folder contains routines for extracting features from raw time series. Run this script interactively in e.g. RStudio (recommended) to produce two processed datasets: trainingDataFeatures.csv and testDataFeatures.csv saved in data folder. The produced datasets are each of size approximately 1.4 GB. It takes in total about 4h to process both training and test set.

Model estimation

The final submission was composed of the output from three modifications of Random Forest classifier. The models are specified in models.R in scripts folder. It takes about 3.5h to train the models.

Generating submission

Along with estimation of the classifiers, in models.R script, the models' predictions are generated and saved in submissions folder. Script majority_voting.R can be used to blend the three submissions by weighted majority voting. This produces the final submission which yielded the best score during the competition.

Final word & Acknowledgements

AAIA`15 Data Mining Contest was an exciting event. Thanks to all the participants for great work and the competition! I hope that the model within this repo will serve as a benchmark for even better performing models. I would also like to thank Marek Gagolewski for inspiring discussions on the competition.

References

  • AAIA'15 Data Mining Competition: Tagging Firefighter Activities at a Fire Scene, URL: https://knowledgepit.fedcsis.org/contest/view.php?id=106 (last access date 13 December 2015).
  • L. Breiman (2001). Random forests. Machine Learning, 45, pp. 5–32.
  • J. Lasek and M. Gagolewski (2015). The Winning Solution to the AAIA’15 Data Mining Competition: Tagging Firefighter Activities at a Fire Scene. Proceedings of the Federated Conference on Computer Science and Information Systems 2015, pp. 375-380, https://fedcsis.org/proceedings/2015/pliks/418.pdf
  • A. Liaw and M. Wiener (2002). Classification and Regression by randomForest. R News 2(3), pp. 18-22.
  • M. Meina, A. Janusz, D. Ślęzak, B. Celmer and A. Karasuski. Tagging Firefighter Activities at the Emergency Scene: Summary of AAIA’15 Data Mining Competition at Knowledge Pit (2015). Proceedings of the Federated Conference on Computer Science and Information Systems 2015, pp. 367–373, https://fedcsis.org/proceedings/2015/pliks/426.pdf