Question on continuous values

[gretelai]

Question on continuous values

[tmontana]

Thanks for the detailed explanation. The precision issue makes sense as well.

[johntmyers]

@tmontana when you say continuous do you mean monotonic (like increasing / decreasing values i.e. timestamps, counters, etc) or values that exist in set ranges like an age, etc.?

[tmontana]

I mean a column containing any real valued number: 0.99836655477 as a random float example.

[amysteier]

During training, the fields in a record of input data are all first parsed into a series of tokens. The parsing is somewhat agnostic as to whether a particular field value is numeric or categorical. Tokens can be either character n-grams or word/number n-grams. A number like 352.4 may end up being one token (352.4), or two tokens (352 and .4) or potentially even more tokens. The chance of a string of characters becoming a token has to do with the probability of that string occurring in the training set.
When generating, the next token is always predicted based on the previous tokens predicted by the model. One of the tokens that can be predicted is the field delimiter (usually something like ","). Again, the model is agnostic as to whether the combined tokens between two field delimiters is numeric or not. The process works quite well for continuous values. It synthesizes an all new series of values for a continuous field, with the same distribution and range as the original series, yet not exactly the same numbers.

[amysteier]

Additionally, it's worth noting that if your float values have an extremely high precision (as in your example), and all you really care about is a precision like .9983, then you're better off first truncating your floats before you do the training. The model has an easier learning task and will end up doing a better job when synthesizing.