What is the best method for classification of time series data? Should I use LSTM or a different method?



I am trying to classify raw accelerometer data x,y,z to its corresponding label.

What is the best architecture for best results?

Or, does anyone have any suggestions on LSTM architectures built on keras with input and output nodes?


Posted 2018-05-29T19:39:32.610

Reputation: 143



I would not use the word "best" but LSTM-RNN are very powerful when dealing with timeseries, simply because they can store information about previous values and exploit the time dependencies between the samples. That said, it is definitely worth going for it.

It has been proven that their performance can be boosted significantly if they are combined with a Convolutional Neural Network (CNN) that can learn the spatial structures in your data, which in this case is one-dimensional.

Take a look at this state-of-the-art method that combines LSTM and CNN, published in 2019.


Posted 2018-05-29T19:39:32.610

Reputation: 3 275


Gated recurrent units (GRUs) is an advanced version of RNN and LSTM. It has less time complexity and good prediction results than LSTM ann RNN. You can use it for dealing with time series dataset. You can boosted it performance by combining advanced version 1D CNN like ResNet, etc.

Faisal shahzad

Posted 2018-05-29T19:39:32.610

Reputation: 1


using lstm or rnn's for time series data is like using a hammer to swat a fly. have you tried time series modeling using classical stat techniques ARCH, ARIMA etc ? the issue of using individual number as inputs (which is what your speedometer is going to give you) means that the states in each lstm / gru cell or unit will have like a 1x1 matrix , meaning 1 parameter. The reason lstm's have been so good with languages is that languages have words / chars and each of them have rich emeddings / encoding schema's that allow the weight matrics inside the lstm cells to have sufficiently large number of params allowing for better generalization. In your case, once u pass single digits of a sequence (i dont know your data so maybe you do have some form of encoding), your model will most likely overfit and perform poorly on real world data. Stat models on the other hand work on generalizing co-effecients for different variables in your system and are, in my view, far better BUT if you do have other features that you can pass and have a rich enough contextual representation of your input data, then nothing comes close to lstm's and even transformers

Vikram Murthy

Posted 2018-05-29T19:39:32.610

Reputation: 178