# Encoding entities with features of continuous values

Given a set of entities, I would like to predict the next in the sequence; for this purpose, I would like to use RNN. However, my first challenge is how to model the entities.

A possible input sequence can be:

EntityType_1 -> \
EntityType_2 -> \
EntityType_1 -> \
EntityType_3


Where each entity, given its type, has a unique set of continuous properties. For instance, the above sequence including features can be:

EntityType_1 (x=0.1, y=0.8) -> \
EntityType_2 (z=0.5) -> \
EntityType_1 (x=0.9, y=0.3) -> \
EntityType_3 (i= 0.2, j=0.7)


I have more than 10,000 entity types, where each has anywhere between 0 to 10s of features.

Any thoughts on related work on modeling/encoding such entities?

• We can use a indicator variable to give a hint to the model whether or not to use the given feature for computation. For instance, in EntityType_2, we don't have features x and y, so we can set their respective indicator variable values to 0. – Shubham Panchal May 5 at 6:39
• I guess that will be most useful when you have a "limited" set of parameters. In my case, the set of parameters is so big, e.g., 8n where n is the number of entity types (note that n is easily bigger than 10,000). – Hamed May 18 at 0:31