Hands-one assistance in Embedding and logisitic regression over aggregated data

[borisRa]

Hi,

I need assistance in three issues :

1. How to I apply the embedding only on the categorical features (I have also continuous )?
2. How do address the following issue with Skflow : [http://stackoverflow.com/questions/33871615/train-a-model-with-probability-response-or-number-of-successes-failures-rather]
3. How do I add the probability estimation for a success in the logistic output ?

Thanks,
Boris

[ilblackdragon]

1. Currently it's not very convenient - I'm working on API making it better.
   To do it - you need to pass everything as continuous matrix and then split it.
   e.g.

def my_model(X, y):
    # X - is [batch_size, n_features], where features split into n_cat + n_cont
    Xcat = tf.cast(tf.slice(X, [0, 0], [X.get_shape()[0], n_cat]), np.int64)
    Xcont = tf.slice(X, [0, n_cat], X.get_shape())

This way Xcat can be passed into categorical_variable and then combined with continues features.

Stay tuned for a better way to do it!

1. @terrytangyuan responded on stackoverflow.
2. Do you mean how to get probability out of the estimator for logistic output? You can just run estimator.predict_proba which will return probabilities per class instead of predicted class.

Let me know if this responds your questions!

[borisRa]

Thanks for the quick response !

1. About the first one : how to combine (column bind for tf object) Xcat & Xcont back to X.
   To apply the deep learning models on X?
2. I meant how to input aggregated data into the logistic regression.Instead of '1' for success and '0' for failure. input the Y attribute total number of successes and failures per aggregation level. For now there is no such support in Scikit => (Logistic regression with a probability response or with number of successes/failures) rather than binary outputs scikit-learn/scikit-learn#6496 (comment))

Is there a solution for this problem in Skflow ?

Thanks again !
Boris

[ilblackdragon]

FeatureColumns are the way to do this now. Please use recent version for Tensorflow to do this.
Thanks!