Featuretools Requests

Sometimes, inference for only a subset of instances in an EntitySet is desirable. Instead of rebuilding an EntitySet to only contain the instances needed, it would also be great to be able to take an existing, larger EntitySet and filter out unneeded instances.

hi, thanks for your suppling featuretools. Now, I want use featuretools to do feature-select, however, I don't know how to use it. I find that API"remove_low_information_features" can be used for Feature Selection. Now, I can't find some example of using this API and the detail about it's principle.

I have daily data of user, and I need to calculate features like first 7 days, second 7 days or last 15 days of the month. The data contains many year so I need to use training_window to calculate for each month. However, the training window doesn't have month unit, only days so I can't do it (since the months have different days) I hope feature labs could consider my idea to update month unit for training window. Thank so much!

It would be helpful, albeit computationally expensive, to be able to perform cardinality analysis on columns to identify everything from indices to categories where the classifier distribution is very unbalanced. If one added in the insane capability of cartesian analysis, this could help bring data in from large old highly abnormalized databases by discovering index column relationships.

I am an AML consultant who is eager to learn Featuretools. I am very much interested in how features are created for transaction monitoring rules using Featuretools. I went through the documentation and some examples on Kaggle; however, I was not able to find a way to create some domain specific features. Let's assume there might be couple of transaction types(cash, wire, check, etc.) and transactions could either be deposit or withdrawal. Feature example could be number of days(within the training window) with daily aggregate total amount of cash deposit transactions is greater than $10K. Another example could be percentage of incoming transaction total divided by outgoing transaction total over 3 day sliding window, assuming beginning and ending of the sliding window changes for everyday within the training window Would you be kind to provide me some examples or codes for creating such features?

For high frequency, high resolution temporal data, it maybe beneficial to create aggregate feature values for different intervals and create a new data frame with these aggregated statistical values replacing the original column. This new data frame can then be used for different machine learning tasks. This could be very useful in IOT data, where sensor observations are often high frequency and have noise in them. A classical way is to apply ``mean'' primitive for each user specified interval of time for a numeric column and create a time series. Since mean itself is a statistical feature, this can be considered as feature time series. If this can be made possible via a simple ft.dfs call, it could prove very useful.

Is it possible to have calculate_feature_matrix to fail quietly? I'm thinking something similar to how many R functions have na.option=pass. The reason for this is when you are scoring on new data, say for production and you have some weird data in it, you may want a warning to come up, but you don't want the whole process to stop. Basically if it were to error out, instead it will fill in the column with na and continue on. Originally posted here: https://github.com/Featuretools/featuretools/issues/36

For many datasets there is only a single table of data. There also might not be a time column. It would be useful if Featuretools had an API to easily perform feature engineering on a single table without having to load it into an EntitySet.

Users often run Featuretools in environments where they either cannot persist data or only have access to a shared file system like S3. For example, a user a using jupyter notebook service like AWS SageMaker would lose access to any saved feature list if they saved their feature definitions to disk. It would be very useful to be able to easily save and load feature definitions from S3.

One functionality that could be very useful in several domains is to be able to calculate a feature for multiple windows in the past. From example, for a customer calculate the "number of purchases made" in past week, past month, past 6 months and year. The reference time point is the cut-off time point. One possible solution could be if ft.dfs can take multiple/an array of training_windows as input. This was also requested by folks at homecredit.eu