The Winning Solution to the AAIA'15 Data Mining Competition: Tagging Firefighter Activities at a Fire Scene

Abstract

Abstract-Multi-sensor based classification of professionals' activities plays a key role in ensuring the success of an his/her goals. In this paper we present the winning solution to the AAIA'15 Tagging Firefighter Activities at a Fire Scene data mining competition. The approach is based on a Random Forest classifier trained on an input data set with almost 5000 features describing the underlying time series of sensory data.

Tagging Firefighter Activities at the Emergency Scene: Summary of AAIA’15 Data Mining Competition at Knowledge Pit

“…They not only provided an insightful view on the state-of-the-art in multidimensional time series analysis but also contained inspiring new ideas, design specifically for the considered problem. The most interesting of these ideas are described by their authors in separate papers submitted for the competition track of AAIA'15 conference [2], [11], [15], [17], [19], [20].…”

“…Some of the teams constructed two independent classifiers, whereas the others merged the two class labels into a single one and trained a single model for a multi-class prediction problem. However, the best performing classification model was firstly trained to predict the first label (the posture) and then, the obtained prediction was used as a new feature for prediction of the second label [15].…”

Automatic Feature Engineering for Prediction of Dangerous Seismic Activities in Coal Mines

“…A recent data mining competition for posture recognition of firefighters [2] inspired different feature engineering approaches that are very effective [3,4,5]. Using the proposed approaches there, from each series of readings the system generates the following types of features: basic statistics (minimum, maximum, range, arithmetic mean, harmonic mean, geometric mean, median, mode, standard deviation, variance, skewness, kurtosis, signal-to-noise ratio, energy, etc.…”

“…); curve fitting parameters [4]; equal-width histogram features [5]; percentile based features (first quartile, median, third quartile, inter-quartile range, amplitude, etc.) [3]; auto-correlation of the signal with several types of correlations (signal processing auto-correlation and Pearson, Spearman and Kendall correlation) [4]; and intercorrelations between each pair of raw time series values using the aforementioned types of correlation coefficients.…”

“…Authors of [3,4,5] demonstrated that this approach can further enhance the predictive performance of the system. Therefore, the framework uses the following time series generators (TSG): first derivatives of the original series [5]; amplitudes, frequencies and magnitudes obtained with fast Fourier transformation [3]; delta series (the deviations of the original values from the mean of the particular block of the time series), which can remove the seasonality in the data; sliding window time series [6]; and combining two time series by multiplying, subtracting or dividing their values [4].…”