# How to predict data from sequence of sequences of variable size?

input data

~~~~SEQ OF LEN 5 DELIM~~~~
1.050
0.275
3.295 0.080
1.910
0.001
~~~~SEQ OF LEN 5 DELIM~~~~
0.034
0.001 0.005 0.167 3.940
0.002 0.014
0.490 3.385 0.000 7.196 0.007
<EMPTY SEQUENCE> (Stackoverflow does not like empty rows)
~~~~SEQ OF LEN 5 DELIM~~~~
0.506 0.323 0.062 0.034 0.429
0.050 0.176
1.376 0.220 0.007 0.369 0.076
1.551 0.000 0.453 0.032 1.156
3.607 0.153 1.607 0.002 0.005
~~~~SEQ OF LEN 5 DELIM~~~~


output data (has physical meaning, not a class number)

10.0
4.0
11.0


Row sequence can be of any length from 0 to 100000. Values can only be > 0.

What I can think of:

1. pad every row sequence up to fixed length, cut rows longer than that;
2. join all rows in a sequence with e.g. -1 value, then pad/cut to fixed length
3. Use one recurrent net for rows, stack its outputs, feed it to another recurrent net.

Any better ideas?

• option 1) sounds more robust, but it depends on the problem at hand. machinelearningmastery.com/… Dec 11 '21 at 17:35
• You are trying to predict. That has several different meanings. If you are looking at waves on the ocean, there is a "local" phenomena where two waves in sequence are very similar but those 200 waves apart are different. There is a consistent physics of wind on the water that makes many waves and the special case where a passing boat makes one or two. What are the "physics" defining the auto-regressive behavior? What are the characteristic scales that are analogous to storms or tides in terms of changes in behavior? Shift, day of week, machine reset, new material lot, and such. Dec 11 '21 at 18:50