can it be used for for multi variate time series

[Sandy4321]

can it be used for for multi variate time series for both categorical and continues features ? pip install flaml[forecast]

[sonichi]

Not as of now. @int-chaos Could you follow up with @Sandy4321 on this issue?

[int-chaos]

Currently, FLAML only supports single time-dependent variable for time series forecasting, as the learners only support univariate time series and do not take in exogenous regressors.

[ghost]

A little off-topic, but I think relevant. For Time Series data I would expect something more intuitive as arguments like ts_forecast -- just forecast means nothing ( not a good naming convention in my opinion).

[sonichi]

@Sandy4321 BTW, is there an existing learner you use for multi variate time series? flaml supports tuning custom learners.

[int-chaos]

At the moment, we will be improving time series forecasting to support continuous features. In regards to categorical features, we ask that users treat it as hierarchical time series forecasting, where the original dataset will be split based on desired categories and treat each as a different forecasting problem.

[sonichi]

@Sandy4321 For the use case you are asking, what do the categorical features mean? Depending on that there are different ways of handling them. Could you provide one example?

[Sandy4321]

sure great code thanks may you clarify : will it work for multivariate time series prediction both regression and classification

each row means the time series, and columns represent different continues or/and discrete/category observation.

1 where all values are continues values weight height age target time 1| 56 160 34 1.2 time 2| 77 170 54 3.5 time 3| 87 167 43 0.7 time 4| 55 198 72 0.5 time 5| 88 176 32 2.3 etc

2 or even will it work for multivariate time series where values are mixture of continues and categorical values for example 2 dimensions have continues values and 3 dimensions are categorical values

        color        weight     gender  height  age      target 

time 1| black 56 m 160 34 yes time 2| white 77 f 170 54 no time 3| yellow 87 m 167 43 yes time 4| white 55 m 198 72 no time 5| white 88 f 176 32 yes

etc

[sonichi]

Thanks @Sandy4321 for the clarification. These two cases will be supported by #254. After it's merged, it'll be great to give it a try with you use case and let us know how it works.

[int-chaos]

The current PR will only support multivariate time series predictions for regression problems (case 1), but the columns can contain different continuous and/or discrete/category observation. In case two, the target is a categorical value ('yes' or 'no') which is a classification problem and is not supported by the the current forecasting learners. Case two will be worked on in a future PR.

[sonichi]

@Sandy4321 v0.7.0 supports use case 1. Could you try it and let us know how it works?

[Sandy4321]

@sonichi your wrote @Sandy4321 v0.7.0 supports use case 1. Could you try it and let us know how it works? may you help to find code example to show how it works pls

[sonichi]

@sonichi your wrote @Sandy4321 v0.7.0 supports use case 1. Could you try it and let us know how it works? may you help to find code example to show how it works pls

Have you checked https://microsoft.github.io/FLAML/docs/Examples/AutoML-Time%20series%20forecast#multivariate-time-series-forecasting-with-exogenous-variables ?