I have a large data set (~ 30 million data points with 5 functions), which I reduced using K-tools to 200,000 clusters. The data are time series with ~ 150,000 time steps. The data on which I would like to train the model is the presence of separate clusters at each time step. The purpose of the predictive model is to generate a generalized sequence similar to the generation of syntactically correct sentences from a model trained in word sequences. The easiest way to think about this data is to try to predict the pixels in the next video frame from the pixels in the current video frame in order to generate a new sequence of frames that come close to the original sequence.
The raw and sparse representation at each time step will be 200,000 binary values representing which clusters are present or not at that time step. Please note that no more than 200 clusters can be present at any one time step, and therefore this representation is extremely rare.
What is the best representation for converting this sparse vector to a dense vector that would be more suitable for time series prediction using Tensorflow?
Initially, I meant RNN / LSTM trained on vectors at each time step, but because of the size of the training vector, I now wonder if the convolution approach is suitable.
Note that I did not actually use tenorflow outside of some simple tutorials, but previously used OpenCV ML features. Please consider me new to your answers.
.