Question: Multivariate time series support

[ghbacct]

LoudML seems to only learn from univariate auto-correlated models, where past values of the target variable are used to predict future values of the same variable.

However I'd like to use LoudML to learn from multivariate time series, where at each time t there are multiple independent variables x_t_1 through x_t_k that are correlated to the target value y_t.

Does LoudML also support learning from multivariate time series data?

For an example of univariate vs multivariate time series see here: https://www.analyticsvidhya.com/blog/2018/09/multivariate-time-series-guide-forecasting-modeling-python-codes/

[ghbacct]

I've just stumbled upon this example: https://loudml.io/guide/en/loudml/reference/1.4/times-dsl-multiple-dimensions.html and I believe it would perform multivariate time series forecasting.

Does anyone have any implementation details on how it works [model used, things to keep in mind, etc.]?

[regel]

@sorinmuchi Hi, mv was dropped in 1.5 (the current master branch). We can re-introduce multivariate feature support in a future release.

What type of multivariate data problem do you have in mind?

[ghbacct]

@regel Sad to hear it was dropped. Is is it a lot of work to add it back? Maybe I can try and help.

As an example let's say I'm trying to replicate the forecasting part of the Google experiment here: https://deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-40/

In which case I would have weather data as an [autoregressive] input, a few other [more or less independent] inputs, and I'm trying to predict the value of datacenter temperature, an autoregressive dependent variable.

So we'd have:

• x_t_1 .. x_t_c - autoregressive _ex. outside_temperature, amount_of_sunlight, TOD, DOY, cooling_fan_speed, air_conditioningpower, etc
• x_t_c+1 ... x_t_k - static parameters _ex., datacenter_insulation_factor, datacenter_emptyvolume, etc.
• y_t - target variable _i.e. datacentertemperature

I would probably learn and predict this at different locations inside the datacenter, and across multiple such datacenters.

_Edit: of course I would not try to predict the weather, but instead would use a weather forecast in order to predict future values for the datacenter_temperature variable._

[ghbacct]

@regel any updates on this?

[lephisto]

I would like to see the return of mv as well. Any Idea when we can see it back?

[regel]

@sorinmuchi thanks a lot for the data centre example. Can we find this data set publicly or privately exchange data?

@lephisto Yes it will be back.

So I've been cleaning up a few things. The CLI is moving to its own package and will be available on PyPI. This is WIP but almost done. Then I will fix #22 and #98 will be the next one on my list.

Our current goal is to create well-designed client libraries that consume the server API over HTTP. They will be available at https://github.com/loudml. Help & contributors for Javascript, Go and Ruby wanted!

[ntlex]

What is the ETA on this? Any updates? Also could you perhaps elaborate on what was the problem with multivariate in the first place and needed to be removed? This will of course help us contribute if needed.