I am trying to a build RNN model to forecast daily sales for several different cities and different product segments (categorical features and multiple inputs for each day) along with numerical features such as traffic , temperature etc (single input for each day). How do I go about building a model using these features. I tried to label encode (Labelencoder) the categorical features and tried to successively train/update a RNN model feeding it with the time series of each city's product segment data but I ended up with huge prediction errors. Is there a way to one hot encode the categorical features and if I do use one hot encoding how do I go about it (what will be my output dimensions)
I would recommend using embedding in this case so that each unique category within a categorical feature is represented as a vector of floats. Then, you can concatenate all vectors coming from different categorical features, and append this long vector along with all other numerical feature values at each timestep.
Besides, if your categorical features are time-invariant (meaning not change over time, e.g. gender), then it might have some issues incorporating static feature values along with other numerical values since RNN is trying to capture the temporal information within the given sequence. But, there is a very helpful link regarding how to add non-temporal features into RNN.
Hope this helps.