TensorFlow中的特征选择

Question

In the TensorFlow documention, it is mentioned that "Through dense embeddings, deep models can generalize better and make predictions on feature pairs that were previously unseen in the training data."

How can we use the dense embeddings in the code and get the new features that TensorFlow makes, which is using generalization and memorization?

Or another way of putting it, how to use TensorFlow as a feature selection algorithm?

Source

Answer 1

Tensor flow does both feature selection and feature transformation to select the best model when optimized to generalize against validation set using suitable regularization. However, the selected input features are hard to discover in dense&deep models. The wide models (1 layer) give a sense of which features are most relevant for the problem. These act as logistic regression layer with weights of the edges representing relative importance of features. Abalation is another way to do feature selection by training NN without one evaluation feature and noting the drop in metrics.

Dense embeddings can be viewed as complex transformations for input feature fit to the training examples. They do not generalize well for unseen data in your test set. They are obtained by defining tensors as : tf.contrib.layers.embedding_column