YLearn for Timeseries data

[ctrivino1]

I have a time series dataset and am wondering how I could use YLearn in with my data? For example lets say I have 4 columns: rate1,rate2,rate3,rate4 (which are all floats). and lets say I wanted to find the causal estimator score for how each rate affects my 'rate1' column (my target feature).

After running the why.fit() function like so 'why.fit(train_data,outcome)', I get my treatment float columns, but am unsure how I can see visualizations or find insights into why/how these treatment variables affect my target feature. how can I find the causal effect of my treatment variables on my outcome variable, and create visualizations with meaningful insights? Thanks!

[BochenLv]

Hi,

Thanks for the question. Currently YLearn is not specially designed for time series data, thus it might not be appropriate to directly apply YLearn to time series data.

However, it is still possible to apply YLearn to the example data suggested in your comment. Specifically, when calling why.fit(), it is better to specify the value of treatment, e.g., why.fit(train_data, outcome, treatment=['rate2', 'rate3']). Then you can find the causal effect of treatment rate2 and rate3 on the outcome rate1 by calling why.causal_effect().

[ctrivino1]

Also Kudos for making this library, its very useful!

[BochenLv]

If you have any suggestions or examples then I'm all ears.

It seems that there indeed lacks such python lib. Currently we only support the causal discovery methods for time series data, perhaps we might start working on including more supports of time series data after a comprehensive investigation.

How do I interpret these result on my outcome variable? (For example is this saying that for every unit increase in aveBedrms there is a -.006 decrease in outcome variable? Thanks!

Yes, this interpretation is basically correct. You can also check our doc for a helpful definition of the average treatment effect (ATE) to get a sense about the interpretation of causal effects in your case.

Also Kudos for making this library, its very useful!

Thanks for saying this 😄

[ctrivino1]

Hi!

I apologize for asking so many questions, I deleted my last two posts because I was able to figure them out after looking at your pyscripts. I did have one suggestion/question however:

Is there anyway to know the columns that are associated to their coefficient values when using the code: double_ml_model.yx_model.coef_?
I have a client who is interested in viewing the coefficients with their associated column names. I also think this would be a helpful feature especially when there are multiple treatment columns being fitted in the model.

I also have been able to create a a tutorial .py script for my colleagues who are interested in this library. If you would like to look at the script and potentially put it in your example notebooks, then please let me know.

Thank you again for your time!

[BochenLv]

Hi,

Sorry for the late reply and thanks for your question.

Is there anyway to know the columns that are associated to their coefficient values when using the code: double_ml_model.yx_model.coef_?

In fact, yx_model.coef_ should not be directly used when viewing the causal effect and the corresponding coef_ does not directly relate to the features (also column names) in the given data. This is because the corresponding coef_ are obtained in a "tricky" way from the given data and is only an intermediate step to estimate the final causal effect.

To view the causal effect, you can directly use the function DoubleML.estimate(). Please see this doc if you are interested in how the coef_ is related to the data and causal effect, where we also specify how to use the DoubleML.estimate() function when there are multiple treatment columns:

treat (float or numpy.ndarray, optional, default=None) – In the case of single discrete treatment, treat should be an int or str of one of all possible treatment values which indicates the value of the intended treatment; in the case of multiple discrete treatment, treat should be a list or an ndarray where treat[i] indicates the value of the i-th intended treatment, for example, when there are multiple discrete treatments, array([‘run’, ‘read’]) means that the treat value of the first treatment is taken as ‘run’ and that of the second treatment is taken as ‘read’; in the case of continuous treatment, treat should be a float or an ndarray.

Please also see the example notebook for several cases of applying DoubleML.

---

I also have been able to create a a tutorial .py script for my colleagues who are interested in this library. If you would like to look at the script and potentially put it in your example notebooks, then please let me know.

It would be great to have this tutorial! Could you please let us look at this script first?

Best regards

[ctrivino1]

Hi,

I am a little confused on the _cross_fit function. I'm just a little confused on exactly what is going on, maybe you could help clarify.

I am using this code: 'double_ml_model.fit(data,outcome,treatment=['MedInc', 'HouseAge'],covariate=['AveRooms', 'AveBedrms', 'Population', 'AveOccup', 'Latitude', 'Longitude'])'

I understand there is an x_model being created, but I'm a little confused on what exactly the x_model is. Are individual models for each treatment being created with 'wv' columns as the independent variables? And for each fold is each treatment being predicted.

I'm a little confused because when I print the 'target_converted = target' varible in the first else statement I get these (almost looks like concatenated) values:
[[ 8.3252 41. ]
[ 8.3014 21. ]
[ 7.2574 52. ]
...
[ 1.7 17. ]
[ 1.8672 18. ]
[ 2.3886 16. ]]

I know these values are values form the two treatments variables, so is the model trying to predict both of these values on each fold it is training and testing? does it see these as one value? I am familiar with machine learning, but I'm used to seeing it in the traditional example way that sklearn has them set up. I also am not an expert programmer so I could also be confusing myself. Thanks!

[Elon991613]

Hello, I have a question for you.

In this case, I found that the main focus is on studying the causal effect of each attribute in a certain feature variable (such as card type) on the transaction amount. Can Ylearn explain the causal effect between each feature variable on the transaction amount? Instead of just analyzing one of the feature variables, a more macro approach is used to perform a causal inference on all variables to obtain the causal effect of each feature variable on the result variable.

Thank you for your answer.