The classification of periodic light curves from non-survey optimized observational data through automated extraction of phase-based visual features

We present Random Forest, Support Vector Machine and Feedforward Neural Network models to classify 2519 variable star light curves. These light curves are generated from a reduction of non-survey optimized observational images gathered by wide-field cameras mounted on the Liverpool Telescope. We extract 16 features found to be highly informative in previous studies and achieve an area under the curve of 0.8495 using a feedforward neural network with 50 hidden neurons trained with stratified 10-fold cross-validation with 3 repeats. We propose using an automated visual feature extraction technique by transforming bin-averaged phase-folded light curves into image based representations. This eliminates much of the noise and the missing phase data, due to sampling defects, should have a less destructive effect on these shape features as they still remain at least partially present. There is also no need for feature engineering as the learning algorithms can learn shape features directly from the light curves. We produced a set of scaled images based on a threshold of data points in each pixel. Training on the same feedforward network, we achieve an area under the curve of 0.6348. By introducing the Period and Amplitude as features into this dataset therefore giving meaning to the dimensions of the image we show this improves to 0.7952. Our current models lack translational-invariance and the method may be better suited to specific sub-classification problems common in the variable object hierarchical multi-class problem.