Unsupervised prediction using Deep learning, LSTM, with simple data set

gwal_p

Dear rapidminer users,
I'm currently working on predicting settlement by time with LSTM. I've followed the LSTM's tutorial process 'Regression of airline passengers' and implemented it to my data for prediction and faced some problems.

1. Operator 'TimeSeries to Tensor' does not allow 'Missing values'.
- The point of my project is to predict future settlement. I've used simple Neural Network for my past project before and predicting settlement was possible by adding future 'Date' with corresponding settlement data as 'Missing data' to Neural Network as test data. So, I tried the same method to LSTM and failed. I've only succeeded in supervised learning with LSTM...
-> Is it possible to implement 'Missing values' to Tensor operators? ...and will it help with unsupervised learning? 

2. Is there other ways to predict future values with unsupervised LSTM?
- Sounds like a silly question but I just can't the way to do it. the tutorial process 'Regression of airline passengers' showed only supervised learning results, not future prediction.

I'm very new to data science and don't have much knowledge on coding, so if LSTM prediction via Rapidminer is out of my capacity, I'll have to find other ways for my project. I have attached excel files and images of it for explaining.

Help would be great, Thanks!

jacobcybulski

It is best to deal with missing values in your time series before you turn them into tensors. You cannot use a normal imputation or replacement with average models for time series. Instead, you will need to "create" missing data by interpolation within a time series. You can use 'Equalize Numerical Indexes' or 'Equalize Time Stamps' for this purpose. When it comes to LSTM, we use them primarily for forecasting. However, LSTMs really do not know what is their own purpose, and so you could use them in unsupervised models. For example, you could use LSTM as part of an Autoencoder and as a result have a clustering system or a time-series data compression, etc. Unfortunately, to create a LSTM autoencoder, you would have to use Python or Java as in the Deep Learning extension, you cannot output a tensor from your deep model. May be in the near future!

jacobcybulski

It is best to deal with missing values in your time series before you turn them into tensors. You cannot use a normal imputation or replacement with average models for time series. Instead, you will need to "create" missing data by interpolation within a time series. You can use 'Equalize Numerical Indexes' or 'Equalize Time Stamps' for this purpose. When it comes to LSTM, we use them primarily for forecasting. However, LSTMs really do not know what is their own purpose, and so you could use them in unsupervised models. For example, you could use LSTM as part of an Autoencoder and as a result have a clustering system or a time-series data compression, etc. Unfortunately, to create a LSTM autoencoder, you would have to use Python or Java as in the Deep Learning extension, you cannot output a tensor from your deep model. May be in the near future!