BlackSwan

Being Right when it Really Matters. #imbalancedData #NLP #transferLearning

This is sanitized code (due to NDA) developed in collaboration/for a startup.

What the code does:

Allows you to customize "off the shelf embeddings" via transfer learning training tasks.

The custom_loss function allows you to pass in an array of any size and apply asymetric weights for misclassification. For example: given the weight matrix

INIT_COST_WEIGHTS = np.ones((3,3))
INIT_COST_WEIGHTS[1,0]=5
INIT_COST_WEIGHTS[2,0]=15
INIT_COST_WEIGHTS[2,1]=1

We are applying a 15x penalty to anything that classified to class 0 from true class 2.

In my case:

Word2Vec was customized via a multi-label/multi-class classification problem to predict "tags/topics" in an email corpus.

The customized embeddings was then used to improve perfromance on a different classification task using the same corpus.