# Problem with features in time series classification

For a student project, we got data like this:

• A one dimensional time series
• A classification label
• Two further features/parameters, unrelated to the time series

At the end of the time series there is an event characterized by two features/parameters (in total these two features give 10 different combinations). The time series consists of 5000 values (roughly between -200 and +200) and may be forming some kindle of wave.

The question is now, how to predict/classify, if the result after this event is positive or negative, given the previous time series.

We wanted to try to use a RNN, but we dont know, how to handle these 2 features of the event best, since it is not part of the time series. Simply adding these two parameters as further inputs does not seem to be appropriate as their importance is then underrepresented in respect to mass of input from the time series.

• Please provide an example row of the dataframe - an example datapoint/timeseries. Your description is not clear. – Nikolas Rieble Jan 20 '17 at 21:34
• @NikolasRieble I have edited my answer, I hope it is enough. It is hard to give an example, since that would really be just a time series, and additionally two numbers for the other two values as described. For context: The time series is an ECG of a heart in fibrilation, and the event is a defibirilation attempt with different settings and the label if it was successfull or not. – Luca Thiede Jan 20 '17 at 22:02
• Compute time-series features (mean, std, skew or other meaningful features possibly) and combine with the other two features as input features. Then train, evaluate and report results. If the results are not good, you can use other features and possibly even use frequency domain features (after fourier transformation). – Nikolas Rieble Jan 20 '17 at 22:07
• That was also our first thought, but than we would loose the time information, and the information about the exact shape of the curve, what we think could be relevant. Thats why we wanted to try RNNs, like LSTM or ESN. Just there we dont know how to handle these extra two variables... – Luca Thiede Jan 20 '17 at 22:17