Tabular with Multiple Input Dimensions

Part 1 (2019)
1

Dear community,

My understanding is that Tabular Model works with 2 dimensions (rows + columns) tables as input dataset.
-> What would be the best way apply the model for multiple (more then 2) input dimensions?

Here the details:
My input data is a table (target is Price)

image

But I believe that in my domain, today’s price might be very dependent from the prev days prices.

So I need this to put as input the previous n prices (in this case n = 3)
image

And this for all the types, so I need an other dimension:
image

Therefore the model can predict daily prices for all types using as input all the prev prices for all types.

-> Do you have any nice ideas how to achieve this with Tabular or with other Models?

-> Pandas DataFrame can have only 2 axis, so I am not able to build this 3d table using DataFrames. Any suggestions?

2

You should represent your data as a multivariate time series. These are some repos to deal with TS in fastai:

1 Like
3

Thanks for the input Victor.
I am using fastai v1 at the moment, so if I understood correctly only this package is compatible with fastai v1 at the moment.

which links to this:

Maybe it’s a quite basic question, but since I am quite new with notebooks, do you have any tips on how can I install the package?

I tried this, but it does not work ( I am using Google Collab)

!pip install git+https://github.com/timeseriesAI/timeseriesAI1/tree/master/fastai_timeseries#egg=timeseriesAI

Basically I have an error while importing fastai_timeseries

from fastai_timeseries import *

in this notebook:

timeseriesAI1/01_Intro_to_Time_Series_Classification.ipynb at master · timeseriesAI/timeseriesAI1 · GitHub

Thanks in advance

4

Try to git clone it instead of pip install it. You should also install (this time via pip) pyts and pyunpack in order to make it work.